Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study

Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study LN Sahu, Ratnesh kumar Namdeo, Sanjay Gupta, Poonam Singh This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6989081/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract DeepFakes, which are fake videos, images, and audio clips designed to look real, have become a serious problem on social media. People can easily be misled by this kind of content, which often spreads false information, damages reputations, and tricks businesses as well as ordinary users. Although researchers have developed many tools to detect DeepFakes, most of these tools have been tested only in perfect laboratory conditions. Social media, however, is a much more chaotic environment, full of low-quality, compressed, and constantly reshared content. In this study, we wanted to find out how well these detection tools actually work in real-life social media platforms like Instagram, TikTok, YouTube, and Facebook where people interact every day. We built a large and realistic collection of videos, audio clips, and images directly from these platforms to reflect what users typically experience—blurry visuals, noisy sounds, and heavily compressed files. We tested several popular DeepFake detection models to measure how accurately, quickly, and reliably they can spot fake content in these everyday conditions. Our results show that even the most advanced detection tools lose about fifteen to twenty percent of their accuracy when working with social media content. Some tools, such as LaDeDa, are fast enough to work in real time on mobile phones but cannot catch all DeepFakes. We also explored a real Instagram case where a fake content campaign spread widely, showing that fully automated systems still struggle to catch every piece of manipulated content. Sometimes, human review is still necessary. Overall, this research emphasizes the urgent need for smarter, faster, and more adaptable DeepFake detection systems that can truly handle the way people share and consume information on social media. DeepFake Social Media Detection Benchmarking Real-World Dataset Misinformation Multimodal Detection Human-in-the-Loop Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction In today’s always-connected world, social media platforms like Instagram, Facebook, TikTok, and YouTube have become part of everyday life. People use these spaces to quickly share images, videos, and audio with friends, family, and large online communities. While this instant sharing has made communication faster and easier, it has also brought serious risks, especially with the growing presence of DeepFakes [ 1 ], [ 2 ]. DeepFakes are fake videos, images, or audio clips created or changed using advanced artificial intelligence methods to make it seem like someone said or did something they never actually did [ 2 ], [ 4 ]. These fake media pieces are often so convincing that most people on social media find it extremely difficult, if not impossible, to tell what is real and what is fake [ 1 ], [ 6 ]. Human behaviour on social media is often shaped by fast decisions, emotional responses, and a natural tendency to trust what looks and sounds real [ 8 ], [ 10 ]. This behaviour makes people especially vulnerable to DeepFakes, which can easily blend in with real posts. For example, a fake video of a celebrity or political leader can quickly go viral and reach thousands or even millions of people before anyone gets a chance to verify whether it is real [ 7 ], [ 19 ]. Most DeepFake detection tools are tested in perfect laboratory conditions using high-quality, clearly marked datasets [ 2 ], [ 3 ], but social media content is very different. It is often heavily compressed, low in quality, and filled with visual noise, making DeepFakes much harder to detect in real social media situations [ 4 ], [ 5 ]. This problem is becoming even more urgent because DeepFake technology is becoming cheaper, faster, and easier for anyone to use [ 7 ], [ 9 ]. Today, even people without special skills can create convincing DeepFakes using free or low-cost online tools [ 2 ], [ 5 ]. This easy access is both helpful and harmful. On one hand, it encourages creative uses, but on the other hand, it increases the risk of harmful activities like spreading false information, stealing identities, and cyberbullying [ 10 ], [ 18 ], [ 19 ]. From a psychological point of view, people naturally trust what they see and hear, especially when the content is emotional or dramatic [ 8 ], [ 17 ]. Social media algorithms make this worse by quickly pushing the most exciting and popular content to more users, often without checking whether the content is true [ 19 ], [ 20 ]. Because of this, dangerous DeepFakes can spread quickly, sometimes before automatic detection tools or human moderators can stop them. Although many DeepFake detection systems have been created in recent years, most of them work well only in carefully controlled environments with clean and clear data [ 3 ], [ 4 ], [ 7 ]. When these systems are used on real social media posts, they often fail because the platforms change the quality and format of the files, which makes DeepFakes harder to find [ 5 ], [ 6 ], [ 9 ]. This research looks closely at this real-world problem by testing how well DeepFake detection tools actually perform on social media content. We built a large, diverse dataset directly from Instagram, TikTok, and YouTube to see how these tools handle the messy, fast-moving content that people deal with every day. Through this study, we aim to offer useful insights into how well current DeepFake detection tools can protect social media users from being misled or harmed during their regular online experiences. Our contributions include: Collected a wide-ranging, real-life dataset that includes videos, images, and audio gathered from different social media platforms. Tested and compared several popular DeepFake detection models using this dataset to see how strong and adaptable they are in real social media situations. Shared a detailed case study of a misinformation campaign on Instagram that used DeepFakes, to closely examine how well current detection methods work in real-life cases. Offered practical suggestions for building better detection systems in the future, highlighting the need for methods that combine multiple types of information and include human decision-making at key steps. 2. Literature Review The following table summarizes the key literature reviewed, including the paper titles, primary findings, and their relevance to DeepFake detection on social media: S.No Paper Title Key Findings 1 Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024 [ 1 ] Introduced a comprehensive multi-modal benchmark dataset for real-world DeepFake detection. 2 Deepfake Generation and Detection: A Benchmark and Survey [ 2 ] Provided a detailed comparison of generation and detection methods, emphasizing diffusion-based DeepFakes. 3 CrossDF: Improving Cross-Domain Deepfake Detection with Deep Information Decomposition [ 3 ] Proposed a cross-domain model that improved generalization across datasets. 4 Locate and Verify: A Two-Stream Network for Improved Deepfake Detection [ 4 ] Developed a two-stream architecture enhancing regional artifact detection. 5 Evolving from Single-modal to Multi-modal Facial Deepfake Detection: A Survey [ 5 ] Highlighted the superior performance of multimodal DeepFake detection approaches. 6 Real-Time Deepfake Detection in the Real-World [ 6 ] Introduced LaDeDa, a real-time DeepFake detection system suitable for edge devices. 7 A Contemporary Survey on Deepfake Detection: Datasets, Algorithms, and Challenges [ 7 ] Comprehensive survey focusing on detection algorithms and dataset limitations. 8 Exploring autonomous methods for deepfake detection: A detailed survey on techniques and evaluation [ 8 ] Reviewed autonomous detection strategies, emphasizing performance in uncontrolled environments. 9 Navigating the Soundscape of Deception: A Comprehensive Survey on Audio Deepfake Generation, Detection, and Future Horizons [ 9 ] Provided in-depth analysis of audio DeepFakes and their detection methods. 10 Deepfake video detection: challenges and opportunities [ 10 ] Identified the main technical challenges in DeepFake video detection. 11 Analysing the landscape of Deep Fake Detection: A Survey [ 11 ] Discussed advancements and gaps in DeepFake detection technologies. 12 A Survey on Detecting Deep Fakes Using Advanced AI-Based Approaches [ 12 ] Presented an overview of AI-driven DeepFake detection methodologies. 13 Performance Comparison and Visualization of AI-Generated-Image Detection Methods [ 13 ] Compared the effectiveness of multiple detection algorithms for AI-generated images. 14 Deepfake Generation and Detection: Case Study and Challenges [ 14 ] Offered case studies detailing DeepFake generation and detection challenges. 15 Deepfake Detection: A Systematic Literature Review [ 15 ] Provided a systematic review covering DeepFake detection techniques up to 2022. 16 A study on data augmentation in voice anti-spoofing [ 16 ] Explored data augmentation methods to improve audio DeepFake detection. 17 Fake video detection among secondary school students: An interdisciplinary study [ 17 ] Examined public awareness and the ability of students to detect fake videos. 18 Deepfake Cyberbullying: The Psychological Toll on Students [ 18 ] Studied the impact of DeepFake-driven cyberbullying on student mental health. 19 Jumio 2024 Online Identity Study [ 19 ] Highlighted increasing concerns regarding identity fraud due to DeepFakes. 20 Deepfake Fraud Doubles Down: 49% of Businesses Now Hit by Audio and Video Scams [ 20 ] Reported on the growing prevalence of DeepFake fraud in business sectors. 3. Methodology The following figure illustrate the methodology in involved in this work: 3.1 Dataset Collection We collected a large and varied dataset that includes more than 50 hours of video, 30 hours of audio, and 2,000 still images from popular social media platforms like Instagram, TikTok, YouTube, and Facebook, covering content shared between 2023 and 2025. To make sure our sources were reliable, we used community flags, compared them with known DeepFake databases, and carefully reviewed and labeled the content with the help of expert reviewers. The dataset we created reflects the same variety found in Deepfake-Eval-2024 [ 1 ] and the WildRF section of LaDeDa [ 6 ]. Along with collecting videos, images, and audio, we also carefully recorded important details like when the content was uploaded, how much it was compressed, how people interacted with it, and which social media platforms it came from. These extra pieces of information help us deeply study how fake content spreads and how reliable detection systems are in real social media environments. 3.2 Model Selection The following models were selected for benchmarking: CrossDF [ 3 ]: Optimized for cross-domain generalization. Locate and Verify [ 4 ]: A two-stream architecture enhancing frame-level detection. LaDeDa and Tiny-LaDeDa [ 6 ]: Real-time detectors optimized for mobile deployment. Face X-ray [ 4 ]: Artifact-based detection leveraging pixel inconsistencies. Each model was retrained and fine-tuned on portions of our dataset to ensure equitable comparison. 3.3 Evaluation Metrics We tested the detection models by measuring how accurately they could find DeepFakes and how often they missed or falsely flagged content. Specifically, we looked at precision, recall, AUC (Area Under Curve), mean average precision (mAP), how fast the models worked, how large the models were, and how well they could handle tricky, intentionally confusing changes. To see how these models would perform in the real world, we also tested them on content that had been compressed and distorted using common social media standards [ 2 ], [ 5 ], [ 8 ]. 3.4 Case Study : We studied a large Instagram misinformation campaign that used DeepFake videos of political leaders. These fake videos were shared through a network of connected accounts working together to spread the content widely. In total, we examined around 120 DeepFake posts to see how well detection models could catch them and to understand when human review and decision-making were still necessary to stop the spread. 4. Benchmarking Results The benchmarking results highlight that DeepFake detection models exhibit considerable performance drops when applied to social media datasets compared to controlled environments. The CrossDF model achieved a precision of 71.2%, recall of 66.4%, AUC of 68.2%, and mean average precision (mAP) of 67.5% on social media content, a significant reduction from its controlled benchmark scores. The Locate and Verify model showed a precision of 73.5%, recall of 69.0%, and AUC of 71.0%, reflecting its robustness at the frame level, though still impacted by compression artifacts.The LaDeDa and Tiny-LaDeDa models demonstrated superior inference speeds of 25 fps and 30 fps, respectively, making them viable for real-time applications on mobile and edge devices. However, their detection accuracy suffered, with Tiny-LaDeDa achieving precision of 68.4%, recall of 65.1%, and AUC of 68.9%. Despite smaller model sizes, both struggled against heavily compressed and re-uploaded content. The models' resistance to compression and adversarial perturbations was measured by performance drop rates. On average, detection precision decreased by 15–20% after social media-induced compression. Table 1 DeepFake Detection Benchmarking Results on Social Media Dataset Model Precision Recall AUC mAP Inference Speed Model Size Compression Resistance (Precision Drop) CrossDF 71.2% 66.4% 68.2% 67.5% 10 fps Large -17% Locate and Verify 73.5% 69.0% 71.0% 70.1% 12 fps Medium -15% LaDeDa 70.4% 67.0% 70.4% 69.3% 25 fps Small -18% Tiny-LaDeDa 68.4% 65.1% 68.9% 68.0% 30 fps Very Small -20% 5. Detailed Case Study The Instagram misinformation campaign we studied involved a group of connected accounts working together to spread politically sensitive DeepFake videos that falsely showed public figures. These fake videos quickly gained attention, with some posts reaching over 500,000 views within just 24 hours. Automated detection systems were able to flag about 55% of the DeepFake content. However, a closer manual review and reports from the community helped uncover another 30% of fake videos that the automated systems missed. Still, about 15% of the DeepFake posts went undetected. These posts were especially hard to catch because they had been heavily compressed, were low in resolution, or used advanced generative methods that successfully hid the usual signs of manipulation. When we studied the timing of detections, we found that using multiple types of clues, such as mismatched lip movements and unusual audio patterns, helped improve detection rates over time. Importantly, involving human reviewers, including both experts and everyday users through community reporting, significantly increased the overall success in catching fake content. This showed that fully automated systems are still not reliable enough to handle the complicated, fast-changing nature of social media on their own. Our deeper analysis also found that these misinformation groups actively worked to take advantage of social media algorithms. They shared content designed to quickly get likes, comments, and shares, which helped their posts spread faster and made it harder for platform moderators to remove them quickly. The time between when a fake post was shared and when it was detected turned out to be a key factor in how far the DeepFakes spread and how much influence they had. 6. Discussion The benchmarking results and the Instagram case study from this research clearly show that current DeepFake detection systems face serious challenges when used in real social media situations. Many detection models work very well in controlled testing, but their ability to detect DeepFakes drops sharply when dealing with the messy and unpredictable nature of social media content. On platforms like Instagram, videos and images are often heavily compressed to save storage space and to load faster. This compression, along with added noise, file conversions, and platform-specific changes, weakens the performance of even the best DeepFake detectors [ 1 ], [ 7 ]. These results support what other recent studies have found—there is an urgent need to develop detection tools that can still work effectively after content has been compressed, changed, and reshared in real-life social media environments [ 6 ]. This study also shows why it is so important to move toward detection systems that look at multiple types of clues, not just visuals. Depending only on visual signals is no longer enough, especially when facing modern DeepFakes made using advanced Generative Adversarial Networks (GANs) and diffusion models. These new DeepFakes can create extremely realistic videos and images that leave very few visible signs of being fake [ 1 ], [ 2 ], [ 4 ]. As Chandra et al. [ 1 ] and Pei et al. [ 2 ] explain, DeepFakes today can look so natural in facial expressions, lighting, and skin texture that visual-only detection tools can easily miss them. Adding audio clues gives another important layer of protection. Small audio mistakes, strange timing patterns, or slight voice differences can reveal a DeepFake, even when the visuals look perfect [ 5 ], [ 9 ]. Looking at how things move from frame to frame—such as natural blinking or smooth body movements—also helps catch fakes that might seem convincing in a still image [ 4 ], [ 10 ]. By combining visual, audio, and motion clues, detection systems can become much better at spotting DeepFakes in the complex and fast-moving world of social media. The shift toward multimodal detection is not merely a theoretical recommendation but a practical necessity, as outlined in several comprehensive surveys [ 5 ], [ 7 ], [ 10 ]. Liu et al. [ 5 ] emphasized the superior performance of multimodal detectors in cross-dataset evaluations, while Gong and Li [ 7 ] underscored that future detectors must exploit synergies across modalities to stay ahead of evolving DeepFake generation technologies. Indeed, Yang et al. [ 3 ] and Shuai et al. [ 4 ] have demonstrated that combining visual and temporal streams significantly improves resilience to cross-domain and cross-platform manipulations. In this study’s Instagram case analysis, unimodal detectors failed to maintain performance across compressed and modified posts, whereas systems incorporating multiple modalities showed comparatively better generalization. Another key finding from this study is the essential role of involving humans directly in DeepFake detection and control. Automated systems are important because they can quickly check large amounts of content, but they are not fully reliable in the noisy, fast-changing, and sometimes deliberately misleading world of social media [ 8 ], [ 12 ]. Mistakes in detection—either missing a DeepFake or wrongly flagging real content—can cause serious harm, especially when the fake content targets political leaders, social activists, or sensitive events. Adding human experts to the review process makes detection much more accurate. Humans can understand social, cultural, and ethical details that current artificial intelligence systems cannot [ 6 ], [ 10 ], [ 11 ]. Having humans involved not only helps prevent wrongful content removal but also builds user and public trust in the system. Another powerful strategy is community reporting, where social media users can flag content they believe might be fake. This approach is practical and can grow with the scale of social media. As Rana et al. [ 15 ] and Turós et al. [ 17 ] have pointed out, helping people become more aware and capable of spotting fake media is key to stopping the spread of harmful DeepFakes. When users are actively involved in this process, studies show that DeepFakes are found and removed more quickly, reducing the time they can cause harm [ 17 ], [ 18 ]. Psychological research by Alexander [ 18 ] also shows the deep emotional and reputational harm that DeepFakes can cause, especially in cases of cyberbullying among young people. This makes it even more important to have fast and reliable detection systems that combine both technology and human judgment. Looking ahead, future research should focus on creating smart, fast, and flexible DeepFake detection models that can keep up with the fast-moving nature of social media. DeepFakes often spread at viral speeds, which means we need detection systems that can work in real time or very close to it to reduce harm quickly [ 6 ], [ 8 ]. Traditional server-based systems may not be fast enough for this job, especially when considering privacy rules and the high cost of processing. A more practical solution is to run these detection models directly on edge devices like smartphones and personal computers. This can help achieve real-time detection without relying on slow or privacy-sensitive cloud servers [ 6 ]. However, to make this possible, we need to develop smaller, energy-efficient models that can still perform well within the limits of mobile and home devices. Another promising approach is federated learning, which can help protect user privacy by allowing model training to happen directly on personal devices, instead of sending private data to centralized servers.. Federated learning allows model updates to happen directly on people’s devices without sending private data to a central location. This protects user privacy while still improving the detection model over time [ 6 ], [ 12 ]. Abdulhamed and Hashim [ 12 ] have shown that federated learning can also help build stronger models by using more diverse data from different countries, languages, and cultures, which is critical for detecting new and unexpected types of DeepFakes. At the same time, future detection systems must include explainable artificial intelligence, often called XAI. As these tools begin to play a bigger role in content moderation and even legal decisions, it becomes essential to explain why a video, image, or audio clip was flagged as fake [ 7 ], [ 10 ], [ 13 ]. Park et al. [ 13 ] showed that when we visualize how AI makes its decisions, people find it easier to trust and understand those decisions. Features like heatmaps that show where the AI is looking, confidence scores, and specific alerts for each type of clue can help users, moderators, and policymakers clearly understand why a piece of content was identified as a DeepFake. This can build trust and help reduce conflicts over content removal. It is also important to remember that the battle between DeepFake creators and detection systems is getting tougher. DeepFake creators are using new tricks to beat detection tools, which means we need to develop smarter and more resilient models that can quickly adapt [ 2 ], [ 10 ]. Pei et al. [ 2 ] and Patel et al. [ 14 ] have discussed how DeepFake technology keeps getting better, using complex anti-detection methods that make the job even harder for researchers. Detection systems must constantly improve to keep up with these advances. This is not just a technical challenge—it is already affecting businesses and the financial world. Reports by Regula [ 20 ] and Jumio [ 19 ] show that DeepFake scams are increasing, with nearly half of surveyed companies facing fake audio and video fraud attempts. These threats highlight the urgent need for strong detection tools that can protect not only individual users but also companies and their reputations. In summary, this study shows that while DeepFake detection systems have made good progress, they are still not ready to fully handle the messy, fast-changing world of social media platforms like Instagram. Moving forward, we need to combine several important strategies: using detection systems that check multiple types of clues, involving human reviewers, building models that work on personal devices, using federated learning to protect privacy, and making detection decisions explainable and easy to understand. Along with better public awareness and stronger community reporting, this multi-layered approach will be key to building a strong defense against the growing threat of DeepFakes in today’s digital society. 7. Conclusion Benchmarking DeepFake detection on social media reveals substantial challenges in bridging the gap between controlled evaluation environments and the complex realities of platform-specific content. The study demonstrates that current detection models, while effective on clean and curated datasets, suffer significant performance degradation when applied to real-world scenarios. Social media platforms introduce compression artifacts, noise, and varying frame rates, all of which mask the subtle visual and audio cues typically exploited by detection systems. Moreover, the rapid evolution of DeepFake generation techniques, particularly those employing advanced GANs and diffusion models, further complicates reliable detection, as these new methods increasingly minimize detectable artifacts. This research highlights the critical importance of multimodal detection approaches that integrate visual, audio, and temporal signals to enhance system robustness. Relying solely on visual features is no longer sufficient given the growing sophistication of DeepFake content. Real-time processing capabilities are also essential, enabling prompt detection and intervention to prevent the viral spread of deceptive media. Additionally, human-in-the-loop strategies, including expert reviews and community-driven reporting mechanisms, significantly bolster the reliability of detection workflows and help mitigate both false positives and false negatives. The introduction of a diverse, real-world DeepFake dataset in this study contributes a valuable benchmark that better represents the complexities of social media environments. This dataset serves as a critical tool for advancing the development and evaluation of next-generation detection systems. Future work should prioritize continuous dataset updates, cross-platform performance assessments, and the incorporation of federated learning and explainable AI to support scalable, privacy-preserving detection solutions. Finally, there is an urgent need to establish industry-wide standards and collaborative frameworks to ensure effective, transparent, and ethical DeepFake detection and moderation across social media platforms. Declarations I hereby declare that this research work titled "Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study" is the result of our independent investigation and original contribution. All sources of information, data, and literature used in this research have been duly acknowledged and referenced. This work has not been submitted, either wholly or in part, for the award of any degree, diploma, or other qualification at any other institution. I affirm that the data collected and the analysis presented in this study are genuine and have been conducted with academic integrity. I understand that any form of plagiarism or academic dishonesty is a serious offense and may result in disciplinary action according to the regulations of the institution. I hereby give my full consent for the publication of this research work titled "Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study" in relevant journals. The contents regarding the publication are listed below: Acknowledgments I would like to express my sincere gratitude to all authors who supported and guided me throughout the course of this research paper. I am especially thankful to my academic mentors and supervisors for their valuable feedback, insightful suggestions, and continuous encouragement that greatly enriched the quality of this work. Funding: NOT APPLICABLE Clinical trial number: NOT APPLICABLE Author contributions: Dr. LN Sahu planned the whole research idea and gave the main direction for this work. He carefully guided the team, shared his valuable advice, and checked the final paper to make sure everything was correct and complete. Dr. Ratnesh Kumar Namdeo collected the data from social media, organized it properly, and set up all the experiments needed to test the DeepFake detection methods. Dr.Sanjay Gupta worked on building and running the DeepFake detection models. He also carefully analyzed the results and prepared the charts and graphs shown in the paper. Dr. Poonam Singh Singh read many past research papers to understand the topic well. She helped in explaining the results clearly and wrote major parts of this paper. All authors equally participate in this research paper. Competing interests: not applicable Ethics Approval This research paper titled "Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study" has been conducted in accordance with established ethical guidelines for academic research. The study did not involve direct interaction with human participants or the collection of personal, sensitive, or identifiable information.The data used in this research were obtained from publicly available sources and social media platforms, ensuring compliance with data privacy regulations and platform-specific terms of use. Where required, permissions for data usage have been duly acknowledged.As this study primarily involved the analysis of publicly accessible content for research purposes, formal approval from an institutional ethics review board was not required. However, all efforts have been made to maintain academic integrity, respect intellectual property rights, and ensure that no harm was caused to individuals or organizations through this research. Consent to Participate As this study titled "Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study" was based on the analysis of publicly available data and did not involve direct interaction with human participants, the requirement for individual consent to participate was not applicable. The research was conducted using open-source datasets and publicly accessible social media content, ensuring that no personal, private, or sensitive information was collected from identifiable individuals. All data used in this study were handled with strict attention to ethical standards and privacy considerations. In cases where third-party data or content was utilized, proper permissions were obtained or the data was used in compliance with platform policies and applicable guidelines. Consent to Publish I hereby give my full consent for the publication of this research work titled "Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study" in relevant academic journals, conference proceedings, institutional repositories, or any other recognized scientific platforms as deemed appropriate. I affirm that all the data, analysis, figures, and content presented in this study are my original work or have been used with proper permissions and acknowledgments where required. I confirm that this work does not infringe upon the rights of any individual, organization, or third party. I understand and agree that upon publication, this work will be publicly accessible for academic, educational, and research purposes, contributing to the broader scientific community. Data Availability The datasets used and analyzed in this study titled " Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study " are obtained from publicly available sources and social media platforms. All relevant data have been properly referenced within the paper. The data supporting the findings of this study are available from the corresponding author upon reasonable request, subject to any applicable terms of use or platform restrictions. No proprietary, confidential, or personally identifiable information has been used in this research. If any additional datasets are required for further investigation, the sources and methods of data collection have been clearly outlined to facilitate replication or extended study. Code Availability The code developed and used for the analysis, benchmarking, and experimental procedures in this study titled " Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study " is available from the corresponding author upon reasonable request. The algorithms and scripts were specifically designed for this research and can be shared for academic, educational, or non-commercial purposes, subject to proper citation of this work. Any third-party libraries or tools used in the development process have been duly acknowledged. References Chandra, N. A. (2025). Mar., Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024, arXiv. Pei, G. (2024). May., Deepfake Generation and Detection: A Benchmark and Survey, arXiv. Yang, S. (2023). Sep., CrossDF: Improving Cross-Domain Deepfake Detection with Deep Information Decomposition, arXiv. Shuai, C. (2023). Sep., Locate and Verify: A Two-Stream Network for Improved Deepfake Detection, arXiv. Liu, P., Tao, Q., & Zhou, J. T. (2024). Evolving from Single-modal to Multi-modal Facial Deepfake Detection: A Survey, arXiv, Jun. Cavia, B. (2024). Jun., Real-Time Deepfake Detection in the Real-World, arXiv. Gong, L. Y., & Li, X. J. (2024). A Contemporary Survey on Deepfake Detection: Datasets, Algorithms, and Challenges, Electronics, vol. 13, Mar. Exploring autonomous methods for deepfake detection: A detailed survey on techniques and evaluation, Heliyon, (Feb. 2025). Wani, T. M., et al. (Nov. 2024). Navigating the Soundscape of Deception: A Comprehensive Survey on Audio Deepfake Generation, Detection, and Future Horizons, Found . Trends Priv. Secur. Deepfake video (2024). detection: challenges and opportunities. Artificial Intelligence Review . Vyas, K. (2024). Jan., Analysing the landscape of Deep Fake Detection: A Survey, Int. J. Intell. Syst. Appl. Eng. Abdulhamed, A. A., & Hashim, A. N. (2024). A Survey on Detecting Deep Fakes Using Advanced AI-Based Approaches, Iraqi Journal of Science, Sep. Park, Y., Na, H., & Choi, D. (2024). Performance Comparison and Visualization of AI-Generated-Image Detection Methods. Ieee Access : Practical Innovations, Open Solutions . Patel, Y. (2023). Deepfake Generation and Detection: Case Study and Challenges. Ieee Access : Practical Innovations, Open Solutions , 11. Md, S., Rana (2022). Deepfake Detection: A Systematic Literature Review, IEEE Access. Cohen, A., et al. (Jun. 2022). A study on data augmentation in voice anti-spoofing . Speech Communication. Turós, S., et al. (Sep. 2024). Fake video detection among secondary school students: An interdisciplinary study . Telematics Inf. Reports. Alexander, A., & Deepfake Cyberbullying. (2025). : The Psychological Toll on Students, Clearing House. Jumio (2024). Online Identity Study, Jumio press release, 2024. Deepfake Fraud Doubles Down (Sep. 2024). 49% of Businesses Now Hit by Audio and Video Scams, Regula. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6989081","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":498201954,"identity":"7a383522-52d2-4180-8343-4d12b2e066f9","order_by":0,"name":"LN Sahu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA2UlEQVRIie3PvQrCMBDA8ZOD63LiGlD0FQqCH/gy6dKpvoHULZPUN/AZnMQxEuzUB3AUBAcnRXAUU9G1zSiYPyTckB9cAHy+X+0IMFeI2o6i40YkQCMLSJaE3cmqxWE515Nh22iItikS8v1ymI0YArNfV5FxFkuICkOEzc0kye1iHMeHKhIWdp9IaS5JPyFLBA9cSCrsYud+8nQnGFqCp6lyIOMFSR0pIwlpgNNMMNX9Zci4O95UKntLc7onj7TbCkxevZg9+jOTeN9Vz7/kG17rXvt8Pt9/9gJoQj2qjwHGLgAAAABJRU5ErkJggg==","orcid":"","institution":"Sri Guru Tegh Bahadur Khalsa College","correspondingAuthor":true,"prefix":"","firstName":"LN","middleName":"","lastName":"Sahu","suffix":""},{"id":498201955,"identity":"7368dd3b-9dec-412d-8eff-59a0bcaeb1db","order_by":1,"name":"Ratnesh kumar Namdeo","email":"","orcid":"","institution":"Sri Guru Tegh Bahadur Khalsa College","correspondingAuthor":false,"prefix":"","firstName":"Ratnesh","middleName":"kumar","lastName":"Namdeo","suffix":""},{"id":498201956,"identity":"aff39d06-9fb3-422c-8c0f-4cdd9ebade36","order_by":2,"name":"Sanjay Gupta","email":"","orcid":"","institution":"Sri Guru Tegh Bahadur Khalsa College","correspondingAuthor":false,"prefix":"","firstName":"Sanjay","middleName":"","lastName":"Gupta","suffix":""},{"id":498201957,"identity":"587c2b29-718d-47f0-abd1-1777d9a35221","order_by":3,"name":"Poonam Singh","email":"","orcid":"","institution":"MATS University","correspondingAuthor":false,"prefix":"","firstName":"Poonam","middleName":"","lastName":"Singh","suffix":""}],"badges":[],"createdAt":"2025-06-27 07:38:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6989081/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6989081/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":88793869,"identity":"c650e2c2-b433-4117-9ecd-daf32cfc91c9","added_by":"auto","created_at":"2025-08-11 13:13:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":598757,"visible":true,"origin":"","legend":"\u003cp\u003eMethodology flowchart for DeepFake detection benchmarking on social media.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-6989081/v1/29113ae4149862b882b80335.png"},{"id":88793867,"identity":"069ffa7e-3742-4eb7-9b4d-d76427d34741","added_by":"auto","created_at":"2025-08-11 13:13:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":61972,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of DeepFake dataset sources and content distribution..\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-6989081/v1/bf888cf7a8494ad608797759.png"},{"id":88793870,"identity":"2a2d5a42-c917-426c-bab3-201d1d0b7e06","added_by":"auto","created_at":"2025-08-11 13:13:24","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":43600,"visible":true,"origin":"","legend":"\u003cp\u003eMultimodal DeepFake detection pipeline integrating visual, audio, and temporal analysis.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-6989081/v1/b41da888b2471f5333cc0e8d.png"},{"id":88793871,"identity":"537b5214-8c82-4520-94e0-fe25cef88241","added_by":"auto","created_at":"2025-08-11 13:13:24","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":52297,"visible":true,"origin":"","legend":"\u003cp\u003eDeepFake detection performance across social media platforms and compression levels.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-6989081/v1/1bf8a21471d2993acf829de6.png"},{"id":88793873,"identity":"a7b71791-ffb7-41ef-8f78-609482147a29","added_by":"auto","created_at":"2025-08-11 13:13:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":34509,"visible":true,"origin":"","legend":"\u003cp\u003eHuman-in-the-loop DeepFake detection workflow combining automated and manual review.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-6989081/v1/2b5397e3a4018682720b71f5.png"},{"id":90082323,"identity":"f5f96f6b-1adb-449f-ad0f-19f768178845","added_by":"auto","created_at":"2025-08-28 09:18:05","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1484009,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6989081/v1/73fb3fea-70cf-40ea-97ad-2bfbe4c4575e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eIn today\u0026rsquo;s always-connected world, social media platforms like Instagram, Facebook, TikTok, and YouTube have become part of everyday life. People use these spaces to quickly share images, videos, and audio with friends, family, and large online communities. While this instant sharing has made communication faster and easier, it has also brought serious risks, especially with the growing presence of DeepFakes [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. DeepFakes are fake videos, images, or audio clips created or changed using advanced artificial intelligence methods to make it seem like someone said or did something they never actually did [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. These fake media pieces are often so convincing that most people on social media find it extremely difficult, if not impossible, to tell what is real and what is fake [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eHuman behaviour on social media is often shaped by fast decisions, emotional responses, and a natural tendency to trust what looks and sounds real [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. This behaviour makes people especially vulnerable to DeepFakes, which can easily blend in with real posts. For example, a fake video of a celebrity or political leader can quickly go viral and reach thousands or even millions of people before anyone gets a chance to verify whether it is real [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Most DeepFake detection tools are tested in perfect laboratory conditions using high-quality, clearly marked datasets [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], but social media content is very different. It is often heavily compressed, low in quality, and filled with visual noise, making DeepFakes much harder to detect in real social media situations [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. This problem is becoming even more urgent because DeepFake technology is becoming cheaper, faster, and easier for anyone to use [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Today, even people without special skills can create convincing DeepFakes using free or low-cost online tools [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. This easy access is both helpful and harmful. On one hand, it encourages creative uses, but on the other hand, it increases the risk of harmful activities like spreading false information, stealing identities, and cyberbullying [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. From a psychological point of view, people naturally trust what they see and hear, especially when the content is emotional or dramatic [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Social media algorithms make this worse by quickly pushing the most exciting and popular content to more users, often without checking whether the content is true [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Because of this, dangerous DeepFakes can spread quickly, sometimes before automatic detection tools or human moderators can stop them. Although many DeepFake detection systems have been created in recent years, most of them work well only in carefully controlled environments with clean and clear data [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. When these systems are used on real social media posts, they often fail because the platforms change the quality and format of the files, which makes DeepFakes harder to find [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis research looks closely at this real-world problem by testing how well DeepFake detection tools actually perform on social media content. We built a large, diverse dataset directly from Instagram, TikTok, and YouTube to see how these tools handle the messy, fast-moving content that people deal with every day. Through this study, we aim to offer useful insights into how well current DeepFake detection tools can protect social media users from being misled or harmed during their regular online experiences.\u003c/p\u003e\u003cp\u003eOur contributions include:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eCollected a wide-ranging, real-life dataset that includes videos, images, and audio gathered from different social media platforms.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTested and compared several popular DeepFake detection models using this dataset to see how strong and adaptable they are in real social media situations.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eShared a detailed case study of a misinformation campaign on Instagram that used DeepFakes, to closely examine how well current detection methods work in real-life cases.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eOffered practical suggestions for building better detection systems in the future, highlighting the need for methods that combine multiple types of information and include human decision-making at key steps.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cp\u003eThe following table summarizes the key literature reviewed, including the paper titles, primary findings, and their relevance to DeepFake detection on social media:\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eS.No\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePaper Title\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eKey Findings\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024 [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIntroduced a comprehensive multi-modal benchmark dataset for real-world DeepFake detection.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeepfake Generation and Detection: A Benchmark and Survey [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eProvided a detailed comparison of generation and detection methods, emphasizing diffusion-based DeepFakes.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCrossDF: Improving Cross-Domain Deepfake Detection with Deep Information Decomposition [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eProposed a cross-domain model that improved generalization across datasets.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLocate and Verify: A Two-Stream Network for Improved Deepfake Detection [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDeveloped a two-stream architecture enhancing regional artifact detection.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eEvolving from Single-modal to Multi-modal Facial Deepfake Detection: A Survey [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eHighlighted the superior performance of multimodal DeepFake detection approaches.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eReal-Time Deepfake Detection in the Real-World [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIntroduced LaDeDa, a real-time DeepFake detection system suitable for edge devices.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eA Contemporary Survey on Deepfake Detection: Datasets, Algorithms, and Challenges [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eComprehensive survey focusing on detection algorithms and dataset limitations.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eExploring autonomous methods for deepfake detection: A detailed survey on techniques and evaluation [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eReviewed autonomous detection strategies, emphasizing performance in uncontrolled environments.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNavigating the Soundscape of Deception: A Comprehensive Survey on Audio Deepfake Generation, Detection, and Future Horizons [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eProvided in-depth analysis of audio DeepFakes and their detection methods.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeepfake video detection: challenges and opportunities [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIdentified the main technical challenges in DeepFake video detection.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAnalysing the landscape of Deep Fake Detection: A Survey [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDiscussed advancements and gaps in DeepFake detection technologies.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eA Survey on Detecting Deep Fakes Using Advanced AI-Based Approaches [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePresented an overview of AI-driven DeepFake detection methodologies.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePerformance Comparison and Visualization of AI-Generated-Image Detection Methods [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCompared the effectiveness of multiple detection algorithms for AI-generated images.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeepfake Generation and Detection: Case Study and Challenges [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eOffered case studies detailing DeepFake generation and detection challenges.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeepfake Detection: A Systematic Literature Review [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eProvided a systematic review covering DeepFake detection techniques up to 2022.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eA study on data augmentation in voice anti-spoofing [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eExplored data augmentation methods to improve audio DeepFake detection.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFake video detection among secondary school students: An interdisciplinary study [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eExamined public awareness and the ability of students to detect fake videos.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeepfake Cyberbullying: The Psychological Toll on Students [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eStudied the impact of DeepFake-driven cyberbullying on student mental health.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eJumio 2024 Online Identity Study [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eHighlighted increasing concerns regarding identity fraud due to DeepFakes.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeepfake Fraud Doubles Down: 49% of Businesses Now Hit by Audio and Video Scams [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eReported on the growing prevalence of DeepFake fraud in business sectors.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"3. Methodology","content":"\u003cp\u003eThe following figure illustrate the methodology in involved in this work:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.1 Dataset Collection\u003c/strong\u003e We collected a large and varied dataset that includes more than 50 hours of video, 30 hours of audio, and 2,000 still images from popular social media platforms like Instagram, TikTok, YouTube, and Facebook, covering content shared between 2023 and 2025. To make sure our sources were reliable, we used community flags, compared them with known DeepFake databases, and carefully reviewed and labeled the content with the help of expert reviewers.\u003c/p\u003e\n\u003cp\u003eThe dataset we created reflects the same variety found in Deepfake-Eval-2024 [\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e] and the WildRF section of LaDeDa [\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e]. Along with collecting videos, images, and audio, we also carefully recorded important details like when the content was uploaded, how much it was compressed, how people interacted with it, and which social media platforms it came from. These extra pieces of information help us deeply study how fake content spreads and how reliable detection systems are in real social media environments.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2 Model Selection\u003c/strong\u003e The following models were selected for benchmarking:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eCrossDF [\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e]: Optimized for cross-domain generalization.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eLocate and Verify [\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e]: A two-stream architecture enhancing frame-level detection.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eLaDeDa and Tiny-LaDeDa [\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e]: Real-time detectors optimized for mobile deployment.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eFace X-ray [\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e]: Artifact-based detection leveraging pixel inconsistencies.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eEach model was retrained and fine-tuned on portions of our dataset to ensure equitable comparison.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3 Evaluation Metrics\u003c/strong\u003e We tested the detection models by measuring how accurately they could find DeepFakes and how often they missed or falsely flagged content. Specifically, we looked at precision, recall, AUC (Area Under Curve), mean average precision (mAP), how fast the models worked, how large the models were, and how well they could handle tricky, intentionally confusing changes. To see how these models would perform in the real world, we also tested them on content that had been compressed and distorted using common social media standards [\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4 Case Study\u003c/strong\u003e: We studied a large Instagram misinformation campaign that used DeepFake videos of political leaders. These fake videos were shared through a network of connected accounts working together to spread the content widely. In total, we examined around 120 DeepFake posts to see how well detection models could catch them and to understand when human review and decision-making were still necessary to stop the spread.\u003c/p\u003e"},{"header":"4. Benchmarking Results","content":"\u003cp\u003eThe benchmarking results highlight that DeepFake detection models exhibit considerable performance drops when applied to social media datasets compared to controlled environments. The CrossDF model achieved a precision of 71.2%, recall of 66.4%, AUC of 68.2%, and mean average precision (mAP) of 67.5% on social media content, a significant reduction from its controlled benchmark scores. The Locate and Verify model showed a precision of 73.5%, recall of 69.0%, and AUC of 71.0%, reflecting its robustness at the frame level, though still impacted by compression artifacts.The LaDeDa and Tiny-LaDeDa models demonstrated superior inference speeds of 25 fps and 30 fps, respectively, making them viable for real-time applications on mobile and edge devices. However, their detection accuracy suffered, with Tiny-LaDeDa achieving precision of 68.4%, recall of 65.1%, and AUC of 68.9%. Despite smaller model sizes, both struggled against heavily compressed and re-uploaded content.\u003c/p\u003e\u003cp\u003eThe models' resistance to compression and adversarial perturbations was measured by performance drop rates. On average, detection precision decreased by 15\u0026ndash;20% after social media-induced compression.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDeepFake Detection Benchmarking Results on Social Media Dataset\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003emAP\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eInference Speed\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eModel Size\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eCompression Resistance (Precision Drop)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCrossDF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e71.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e66.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e68.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e67.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e10 fps\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003eLarge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-17%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLocate and Verify\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e73.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e69.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e71.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e70.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e12 fps\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003eMedium\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-15%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLaDeDa\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e70.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e67.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e70.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e69.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e25 fps\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003eSmall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-18%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTiny-LaDeDa\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e68.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e65.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e68.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e68.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e30 fps\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003eVery Small\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-20%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"5. Detailed Case Study","content":"\u003cp\u003eThe Instagram misinformation campaign we studied involved a group of connected accounts working together to spread politically sensitive DeepFake videos that falsely showed public figures. These fake videos quickly gained attention, with some posts reaching over 500,000 views within just 24 hours. Automated detection systems were able to flag about 55% of the DeepFake content. However, a closer manual review and reports from the community helped uncover another 30% of fake videos that the automated systems missed. Still, about 15% of the DeepFake posts went undetected. These posts were especially hard to catch because they had been heavily compressed, were low in resolution, or used advanced generative methods that successfully hid the usual signs of manipulation. When we studied the timing of detections, we found that using multiple types of clues, such as mismatched lip movements and unusual audio patterns, helped improve detection rates over time. Importantly, involving human reviewers, including both experts and everyday users through community reporting, significantly increased the overall success in catching fake content. This showed that fully automated systems are still not reliable enough to handle the complicated, fast-changing nature of social media on their own. Our deeper analysis also found that these misinformation groups actively worked to take advantage of social media algorithms. They shared content designed to quickly get likes, comments, and shares, which helped their posts spread faster and made it harder for platform moderators to remove them quickly. The time between when a fake post was shared and when it was detected turned out to be a key factor in how far the DeepFakes spread and how much influence they had.\u003c/p\u003e"},{"header":"6. Discussion","content":"\u003cp\u003eThe benchmarking results and the Instagram case study from this research clearly show that current DeepFake detection systems face serious challenges when used in real social media situations. Many detection models work very well in controlled testing, but their ability to detect DeepFakes drops sharply when dealing with the messy and unpredictable nature of social media content. On platforms like Instagram, videos and images are often heavily compressed to save storage space and to load faster. This compression, along with added noise, file conversions, and platform-specific changes, weakens the performance of even the best DeepFake detectors [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. These results support what other recent studies have found\u0026mdash;there is an urgent need to develop detection tools that can still work effectively after content has been compressed, changed, and reshared in real-life social media environments [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis study also shows why it is so important to move toward detection systems that look at multiple types of clues, not just visuals. Depending only on visual signals is no longer enough, especially when facing modern DeepFakes made using advanced Generative Adversarial Networks (GANs) and diffusion models. These new DeepFakes can create extremely realistic videos and images that leave very few visible signs of being fake [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. As Chandra et al. [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] and Pei et al. [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] explain, DeepFakes today can look so natural in facial expressions, lighting, and skin texture that visual-only detection tools can easily miss them. Adding audio clues gives another important layer of protection. Small audio mistakes, strange timing patterns, or slight voice differences can reveal a DeepFake, even when the visuals look perfect [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Looking at how things move from frame to frame\u0026mdash;such as natural blinking or smooth body movements\u0026mdash;also helps catch fakes that might seem convincing in a still image [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. By combining visual, audio, and motion clues, detection systems can become much better at spotting DeepFakes in the complex and fast-moving world of social media.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe \u003cb\u003eshift toward multimodal detection\u003c/b\u003e is not merely a theoretical recommendation but a practical necessity, as outlined in several comprehensive surveys [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Liu et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] emphasized the superior performance of multimodal detectors in cross-dataset evaluations, while Gong and Li [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] underscored that future detectors must exploit synergies across modalities to stay ahead of evolving DeepFake generation technologies. Indeed, Yang et al. [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] and Shuai et al. [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] have demonstrated that combining visual and temporal streams significantly improves resilience to cross-domain and cross-platform manipulations. In this study\u0026rsquo;s Instagram case analysis, unimodal detectors failed to maintain performance across compressed and modified posts, whereas systems incorporating multiple modalities showed comparatively better generalization.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAnother key finding from this study is the essential role of involving humans directly in DeepFake detection and control. Automated systems are important because they can quickly check large amounts of content, but they are not fully reliable in the noisy, fast-changing, and sometimes deliberately misleading world of social media [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Mistakes in detection\u0026mdash;either missing a DeepFake or wrongly flagging real content\u0026mdash;can cause serious harm, especially when the fake content targets political leaders, social activists, or sensitive events. Adding human experts to the review process makes detection much more accurate. Humans can understand social, cultural, and ethical details that current artificial intelligence systems cannot [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Having humans involved not only helps prevent wrongful content removal but also builds user and public trust in the system.\u003c/p\u003e\u003cp\u003eAnother powerful strategy is community reporting, where social media users can flag content they believe might be fake. This approach is practical and can grow with the scale of social media. As Rana et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] and Tur\u0026oacute;s et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] have pointed out, helping people become more aware and capable of spotting fake media is key to stopping the spread of harmful DeepFakes. When users are actively involved in this process, studies show that DeepFakes are found and removed more quickly, reducing the time they can cause harm [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Psychological research by Alexander [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] also shows the deep emotional and reputational harm that DeepFakes can cause, especially in cases of cyberbullying among young people. This makes it even more important to have fast and reliable detection systems that combine both technology and human judgment.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eLooking ahead, future research should focus on creating smart, fast, and flexible DeepFake detection models that can keep up with the fast-moving nature of social media. DeepFakes often spread at viral speeds, which means we need detection systems that can work in real time or very close to it to reduce harm quickly [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Traditional server-based systems may not be fast enough for this job, especially when considering privacy rules and the high cost of processing. A more practical solution is to run these detection models directly on edge devices like smartphones and personal computers. This can help achieve real-time detection without relying on slow or privacy-sensitive cloud servers [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, to make this possible, we need to develop smaller, energy-efficient models that can still perform well within the limits of mobile and home devices. Another promising approach is federated learning, which can help protect user privacy by allowing model training to happen directly on personal devices, instead of sending private data to centralized servers.. Federated learning allows model updates to happen directly on people\u0026rsquo;s devices without sending private data to a central location. This protects user privacy while still improving the detection model over time [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Abdulhamed and Hashim [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] have shown that federated learning can also help build stronger models by using more diverse data from different countries, languages, and cultures, which is critical for detecting new and unexpected types of DeepFakes.\u003c/p\u003e\u003cp\u003eAt the same time, future detection systems must include explainable artificial intelligence, often called XAI. As these tools begin to play a bigger role in content moderation and even legal decisions, it becomes essential to explain why a video, image, or audio clip was flagged as fake [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Park et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] showed that when we visualize how AI makes its decisions, people find it easier to trust and understand those decisions. Features like heatmaps that show where the AI is looking, confidence scores, and specific alerts for each type of clue can help users, moderators, and policymakers clearly understand why a piece of content was identified as a DeepFake. This can build trust and help reduce conflicts over content removal. It is also important to remember that the battle between DeepFake creators and detection systems is getting tougher. DeepFake creators are using new tricks to beat detection tools, which means we need to develop smarter and more resilient models that can quickly adapt [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Pei et al. [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] and Patel et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] have discussed how DeepFake technology keeps getting better, using complex anti-detection methods that make the job even harder for researchers. Detection systems must constantly improve to keep up with these advances.\u003c/p\u003e\u003cp\u003eThis is not just a technical challenge\u0026mdash;it is already affecting businesses and the financial world. Reports by Regula [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and Jumio [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] show that DeepFake scams are increasing, with nearly half of surveyed companies facing fake audio and video fraud attempts. These threats highlight the urgent need for strong detection tools that can protect not only individual users but also companies and their reputations. In summary, this study shows that while DeepFake detection systems have made good progress, they are still not ready to fully handle the messy, fast-changing world of social media platforms like Instagram. Moving forward, we need to combine several important strategies: using detection systems that check multiple types of clues, involving human reviewers, building models that work on personal devices, using federated learning to protect privacy, and making detection decisions explainable and easy to understand. Along with better public awareness and stronger community reporting, this multi-layered approach will be key to building a strong defense against the growing threat of DeepFakes in today\u0026rsquo;s digital society.\u003c/p\u003e"},{"header":"7. Conclusion","content":"\u003cp\u003eBenchmarking DeepFake detection on social media reveals substantial challenges in bridging the gap between controlled evaluation environments and the complex realities of platform-specific content. The study demonstrates that current detection models, while effective on clean and curated datasets, suffer significant performance degradation when applied to real-world scenarios. Social media platforms introduce compression artifacts, noise, and varying frame rates, all of which mask the subtle visual and audio cues typically exploited by detection systems. Moreover, the rapid evolution of DeepFake generation techniques, particularly those employing advanced GANs and diffusion models, further complicates reliable detection, as these new methods increasingly minimize detectable artifacts. This research highlights the critical importance of multimodal detection approaches that integrate visual, audio, and temporal signals to enhance system robustness. Relying solely on visual features is no longer sufficient given the growing sophistication of DeepFake content. Real-time processing capabilities are also essential, enabling prompt detection and intervention to prevent the viral spread of deceptive media. Additionally, human-in-the-loop strategies, including expert reviews and community-driven reporting mechanisms, significantly bolster the reliability of detection workflows and help mitigate both false positives and false negatives. The introduction of a diverse, real-world DeepFake dataset in this study contributes a valuable benchmark that better represents the complexities of social media environments. This dataset serves as a critical tool for advancing the development and evaluation of next-generation detection systems. Future work should prioritize continuous dataset updates, cross-platform performance assessments, and the incorporation of federated learning and explainable AI to support scalable, privacy-preserving detection solutions. Finally, there is an urgent need to establish industry-wide standards and collaborative frameworks to ensure effective, transparent, and ethical DeepFake detection and moderation across social media platforms.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eI hereby declare that this research work titled \u003cstrong\u003e\"Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study\"\u003c/strong\u003e is the result of our independent investigation and original contribution. All sources of information, data, and literature used in this research have been duly acknowledged and referenced.\u003c/p\u003e\n\u003cp\u003eThis work has not been submitted, either wholly or in part, for the award of any degree, diploma, or other qualification at any other institution. I affirm that the data collected and the analysis presented in this study are genuine and have been conducted with academic integrity.\u003c/p\u003e\n\u003cp\u003eI understand that any form of plagiarism or academic dishonesty is a serious offense and may result in disciplinary action according to the regulations of the institution. I hereby give my full consent for the publication of this research work titled \u003cstrong\u003e\"Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study\"\u003c/strong\u003e in relevant journals.\u003c/p\u003e\n\u003cp\u003eThe contents regarding the publication are listed below:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u0026nbsp;I\u0026nbsp;would like to express my sincere gratitude to all authors who supported and guided me throughout the course of this research paper. I am especially thankful to my academic mentors and supervisors for their valuable feedback, insightful suggestions, and continuous encouragement that greatly enriched the quality of this work.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e \u003cstrong\u003eNOT APPLICABLE\u003c/strong\u003e\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eClinical trial number:\u003c/strong\u003e \u003cstrong\u003eNOT APPLICABLE\u003c/strong\u003e\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eAuthor contributions:\u0026nbsp;\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eDr. LN Sahu planned the whole research idea and gave the main direction for this work. He carefully guided the team, shared his valuable advice, and checked the final paper to make sure everything was correct and complete. Dr. Ratnesh Kumar Namdeo collected the data from social media, organized it properly, and set up all the experiments needed to test the DeepFake detection methods. Dr.Sanjay Gupta worked on building and running the DeepFake detection models. He also carefully analyzed the results and prepared the charts and graphs shown in the paper. Dr. Poonam Singh Singh read many past research papers to understand the topic well. She helped in explaining the results clearly and wrote major parts of this paper. All authors equally participate in this research paper.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eCompeting interests: \u0026nbsp;not applicable\u003c/strong\u003e\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eEthics Approval\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThis research paper titled \"Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study\" has been conducted in accordance with established ethical guidelines for academic research. The study did not involve direct interaction with human participants or the collection of personal, sensitive, or identifiable information.The data used in this research were obtained from publicly available sources and social media platforms, ensuring compliance with data privacy regulations and platform-specific terms of use. Where required, permissions for data usage have been duly acknowledged.As this study primarily involved the analysis of publicly accessible content for research purposes, formal approval from an institutional ethics review board was not required. However, all efforts have been made to maintain academic integrity, respect intellectual property rights, and ensure that no harm was caused to individuals or organizations through this research.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eConsent to Participate\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eAs this study titled \"Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study\" was based on the analysis of publicly available data and did not involve direct interaction with human participants, the requirement for individual consent to participate was not applicable. The research was conducted using open-source datasets and publicly accessible social media content, ensuring that no personal, private, or sensitive information was collected from identifiable individuals. All data used in this study were handled with strict attention to ethical standards and privacy considerations. In cases where third-party data or content was utilized, proper permissions were obtained or the data was used in compliance with platform policies and applicable guidelines.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eConsent to Publish\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eI hereby give my full consent for the publication of this research work titled \"Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study\" in relevant academic journals, conference proceedings, institutional repositories, or any other recognized scientific platforms as deemed appropriate. I affirm that all the data, analysis, figures, and content presented in this study are my original work or have been used with proper permissions and acknowledgments where required. I confirm that this work does not infringe upon the rights of any individual, organization, or third party. I understand and agree that upon publication, this work will be publicly accessible for academic, educational, and research purposes, contributing to the broader scientific community.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe datasets used and analyzed in this study titled \"\u003cstrong\u003eBenchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study\u003c/strong\u003e\" are obtained from publicly available sources and social media platforms. All relevant data have been properly referenced within the paper. The data supporting the findings of this study are available from the corresponding author upon reasonable request, subject to any applicable terms of use or platform restrictions. No proprietary, confidential, or personally identifiable information has been used in this research. If any additional datasets are required for further investigation, the sources and methods of data collection have been clearly outlined to facilitate replication or extended study.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eCode Availability\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe code developed and used for the analysis, benchmarking, and experimental procedures in this study titled \"\u003cstrong\u003eBenchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study\u003c/strong\u003e\" is available from the corresponding author upon reasonable request. The algorithms and scripts were specifically designed for this research and can be shared for academic, educational, or non-commercial purposes, subject to proper citation of this work. Any third-party libraries or tools used in the development process have been duly acknowledged.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eChandra, N. A. (2025). Mar., Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024, arXiv.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePei, G. (2024). May., Deepfake Generation and Detection: A Benchmark and Survey, arXiv.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang, S. (2023). Sep., CrossDF: Improving Cross-Domain Deepfake Detection with Deep Information Decomposition, arXiv.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShuai, C. (2023). Sep., Locate and Verify: A Two-Stream Network for Improved Deepfake Detection, arXiv.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, P., Tao, Q., \u0026amp; Zhou, J. T. (2024). Evolving from Single-modal to Multi-modal Facial Deepfake Detection: A Survey, arXiv, Jun.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCavia, B. (2024). Jun., Real-Time Deepfake Detection in the Real-World, arXiv.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGong, L. Y., \u0026amp; Li, X. J. (2024). A Contemporary Survey on Deepfake Detection: Datasets, Algorithms, and Challenges, Electronics, vol. 13, Mar.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eExploring autonomous methods for deepfake detection: A detailed survey on techniques and evaluation, Heliyon, (Feb. 2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWani, T. M., et al. (Nov. 2024). \u003cem\u003eNavigating the Soundscape of Deception: A Comprehensive Survey on Audio Deepfake Generation, Detection, and Future Horizons, Found\u003c/em\u003e. Trends Priv. Secur.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeepfake video (2024). detection: challenges and opportunities. \u003cem\u003eArtificial Intelligence Review\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVyas, K. (2024). Jan., Analysing the landscape of Deep Fake Detection: A Survey, Int. J. Intell. Syst. Appl. Eng.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbdulhamed, A. A., \u0026amp; Hashim, A. N. (2024). A Survey on Detecting Deep Fakes Using Advanced AI-Based Approaches, Iraqi Journal of Science, Sep.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePark, Y., Na, H., \u0026amp; Choi, D. (2024). Performance Comparison and Visualization of AI-Generated-Image Detection Methods. \u003cem\u003eIeee Access : Practical Innovations, Open Solutions\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePatel, Y. (2023). Deepfake Generation and Detection: Case Study and Challenges. \u003cem\u003eIeee Access : Practical Innovations, Open Solutions\u003c/em\u003e, 11.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMd, S., Rana (2022). Deepfake Detection: A Systematic Literature Review, IEEE Access.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCohen, A., et al. (Jun. 2022). \u003cem\u003eA study on data augmentation in voice anti-spoofing\u003c/em\u003e. Speech Communication.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTur\u0026oacute;s, S., et al. (Sep. 2024). \u003cem\u003eFake video detection among secondary school students: An interdisciplinary study\u003c/em\u003e. Telematics Inf. Reports.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlexander, A., \u0026amp; Deepfake Cyberbullying. (2025). : The Psychological Toll on Students, Clearing House.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJumio (2024). Online Identity Study, Jumio press release, 2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeepfake Fraud Doubles Down (Sep. 2024). 49% of Businesses Now Hit by Audio and Video Scams, Regula.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"DeepFake, Social Media, Detection Benchmarking, Real-World Dataset, Misinformation, Multimodal Detection, Human-in-the-Loop","lastPublishedDoi":"10.21203/rs.3.rs-6989081/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6989081/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDeepFakes, which are fake videos, images, and audio clips designed to look real, have become a serious problem on social media. People can easily be misled by this kind of content, which often spreads false information, damages reputations, and tricks businesses as well as ordinary users. Although researchers have developed many tools to detect DeepFakes, most of these tools have been tested only in perfect laboratory conditions. Social media, however, is a much more chaotic environment, full of low-quality, compressed, and constantly reshared content. In this study, we wanted to find out how well these detection tools actually work in real-life social media platforms like Instagram, TikTok, YouTube, and Facebook where people interact every day. We built a large and realistic collection of videos, audio clips, and images directly from these platforms to reflect what users typically experience\u0026mdash;blurry visuals, noisy sounds, and heavily compressed files. We tested several popular DeepFake detection models to measure how accurately, quickly, and reliably they can spot fake content in these everyday conditions. Our results show that even the most advanced detection tools lose about fifteen to twenty percent of their accuracy when working with social media content. Some tools, such as LaDeDa, are fast enough to work in real time on mobile phones but cannot catch all DeepFakes. We also explored a real Instagram case where a fake content campaign spread widely, showing that fully automated systems still struggle to catch every piece of manipulated content. Sometimes, human review is still necessary. Overall, this research emphasizes the urgent need for smarter, faster, and more adaptable DeepFake detection systems that can truly handle the way people share and consume information on social media.\u003c/p\u003e","manuscriptTitle":"Benchmarking DeepFake Detection on Social Media: Real-World Dataset and Case Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-11 13:13:19","doi":"10.21203/rs.3.rs-6989081/v1","editorialEvents":[{"type":"communityComments","content":8}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7be34d1b-80db-4aeb-9af2-0276bb458923","owner":[],"postedDate":"August 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-08-28T09:09:49+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-11 13:13:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6989081","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6989081","identity":"rs-6989081","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}