Real-Time Deepfake Detection Using a Hybrid Mobile Net-LSTM Model For Image and Video Analysis

Real-Time Deepfake Detection Using a Hybrid Mobile Net-LSTM Model For Image and Video Analysis | 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 Real-Time Deepfake Detection Using a Hybrid Mobile Net-LSTM Model For Image and Video Analysis S Ambika, B R Sumanth, Yerramsetty Harini This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6530533/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 The rise of deepfake technology has led to an urgent need for robust detection mechanisms to combat misinformation and digital fraud. Deepfake media, generated using deep learning techniques, has become increasingly sophisticated, making it difficult to distinguish between real and manipulated content. This paper presents a hybrid MobileNet-LSTM model designed for real-time deepfake detection in both image and video formats. MobileNet efficiently extracts spatial features, while LSTM captures temporal dependencies, ensuring higher accuracy in detecting manipulated content. The proposed method is evaluated on benchmark datasets, demonstrating improved accuracy, computational efficiency, and resilience against adversarial manipulations. Our results indicate the model's effectiveness in real-time applications, making it a suitable candidate for social media and digital forensic platforms. Furthermore, a comparative study with state of the art models highlights the advantages of our approach in terms of precision, recall, and robustness against adversarial attacks. Deepfake Detection MobileNet LSTM Image Processing Video Analysis Artificial Intelligence Misinformation Real-time Processing Cybersecurity Figures Figure 1 Figure 2 Figure 3 Figure 4 1 Introduction The emergence of artificial intelligence (AI) has given rise to many groundbreaking technologies, one of the most contentious being deepfake technology. Deepfakes, or highly realistic manipulations of visual and audio material, have become popular in recent times because of their potential for misinformation and social manipulation. The word "deepfake" originated from the couple of deep learning algorithms, mainly Generative Adversarial Networks (GANs), and the word "fake." Deepfake technology has facilitated people to produce fake videos, photos, and audio files that cannot be distinguished from authentic content. The doctored media are usually applied for nefarious purposes, like circulating false information, conducting political propaganda, or even cyberbullying and slander. As deepfake technology continues to advance and become more accessible, the detection of such manipulated media has become an urgent concern in numerous industries, ranging from news to social media, entertainment, and law enforcement. Deepfake detection tools generally used either Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) to process static images and dynamic video content. CNNs are good at extracting spatial features from images but can be challenged by videos, where the temporal dimension also needs to be taken into account. RNNs and its more complex variant, LSTM, can scan sequences of video frames to identify very subtle discrepancies that may not be apparent in an individual frame. Existing deepfake detection systems are good for many applications, but they often have issues with real-time detection, scalability, and computational cost. This work suggests a new hybrid solution which combines the capabilities of MobileNet and LSTM networks to overcome these limitations.. 2 Related Work The field of deepfake detection has gained serious consideration amidst the growing misuse of AI-synthesized content. Researchers have tried various methodologies to detect deepfakes in an effective way, ranging from traditional hand crafted approaches to current state of the deep learning architectures. This section of the chapter discusses main contributions in deepfake generation techniques and hybrid approaches. Convolutional neural networks (CNNs) were comprehensively employed for extracting spatial features in deepfake detection. Rossler et al. (2019) proposed XceptionNet and trained it on the FaceForensics + + dataset, and it performed in deepfake video detection (Rossler et al., 2019). Nguyen et al. (2020) used capsule networks to increase generalization capability over various deepfake datasets and overcome the constraints of CNN-based models when dealing with unseen deepfake styles (Nguyen et al., 2020). However, CNNs lack temporal awareness, making them less effective for video based deepfake detection. Apart from enhancing temporal consistency analysis, recurrent neural networks (RNNs) and long short-term memory (LSTM) models have also been incorporated into deep-fake detectors. Guera and Delp (2018) employed the use of an LSTM-based model to analyze frame sequences to advance detection performance of deep-fake videos by catching inter-frame inconsistencies (Guera & Delp, 2018).. Zhang et al. (2021) built on the concept by adding an additional mechanism to LSTM models to enhance resilience against adversarial manipulated deepfake content (Zhang et al., 2021). Hybrid techniques integrating CNNs with temporal models have appeared as effective solutions. Zhao et al. (2022) developed a MobileNet-LSTM architecture, utilizing MobileNet for effective spatial feature extraction and LSTM for sequential learning. Their approach drastically minimized computational expenses while ensuring high detection accuracy (Zhao et al., 2022). Li et al. (2023) presented a transformer-based method that achieved state-of-the-art performance in deepfake detection but needed a large amount of computational power, making real-time deployment more challenging (Li et al., 2023). There has also been research in adversarial robustness and explainability of deepfakes for detection recently. Wang et al. ( 2023 ) investigated the application of adversarial training methods for enhancing the robustness of the model to perturbations applied by deepfake generation models (Wang et al., 2023 ). Explainable AI uses like Grad-CAM heatmaps have been used within deepfake detection pipelines for supplying interpretability into model decision-making (Huang et al., 2023). Together, the studies point out the pros and cons of current deepfake detection methods. Our suggested MobileNet-LSTM framework improves previous work by increasing the balance between computational efficiency and detection precision, ensuring it can perform in real-time scenarios. 3 Research Methodology Deepfake detection model proposed here is a hybrid deep learning-based model using MobileNet for spatial feature extraction and LSTM for temporal sequence analysis. This section explains system architecture, pre-processing of data, model training, and evaluation techniques used in order to achieve high accuracy and computational efficiency. Also, This work solves the problem of identifying deepfakes by creating a hybrid deepfake detection system combining two strong methods: MobileNet for effective spatial feature extraction and Long Short-Term Memory (LSTM) networks for processing temporal dependencies in video data. 3.1 System Acrchitecture The deepfake detection system owns a multi-stage pipeline consisting of preprocessing, feature extraction, sequential analysis, and classification. The major parts of the architecture are: Input Processing : The videos and images are preprocessed by the frame capture at a steady rate and the scaling of them to a constant resolution for consistency. Feature Extraction using MobileNet : MobileNet, a lightweight CNN, extracts spatial features from each frame, capturing essential texture and edge information that helps distinguish real from fake media. Temporal Analysis using LSTM : Extracted spatial features are passed through an LSTM network, which learns sequential dependencies across video frames, detecting inconsistencies that are characteristic of deepfake videos. Classification Layer : The final output is passed through a fully connected softmax layer that classifies media as either real or deepfake. 3.2 Data Preprocessing Data preprocessing is crucial in improving model performance and ensuring robust generalization. The preprocessing steps include: Dataset Selection : The model is trained and tested on benchmark datasets such as FaceForensics++, Celeb-DF, and DFDC. Frame Extraction : Video frames are sampled at a rate of 10 frames per second to ensure sufficient temporal coverage. Normalization and Augmentation : Pixel values are normalized, and data augmentation techniques such as rotation, brightness adjustments, and Gaussian noise are applied to enhance dataset variability. Face Detection and Cropping : OpenCV and MTCNN-based face detection algorithms are used to extract only the facial regions, removing unnecessary background noise. 3.3 Model Training and Optimization The deepfake detection model is trained using supervised learning techniques with the following configurations: Loss Function : Categorical cross-entropy is used to optimize classification accuracy. Optimizer : The Adam optimizer is employed with a learning rate of 0.001 for faster convergence. Batch Size and Epochs : The model is trained for 50 epochs with a batch size of 32. Regularization Techniques : Dropout layers and L2 regularization are incorporated to prevent overfitting. 3.4 Evaluation Metrics The performance of the proposed model is evaluated using the following metrics: Accuracy : Estimates a portion of labeled deep fake and real videos properly. Precision and Recall : The performance of model is estimated to classify deepfake data accurately at the cost of false positives and false negatives. F1-Score : Harmonic mean of precision and recall to yield an even measure of evaluation. ROC-AUC Curve : Assesses the model's ability to differentiate between fake and real media across various classification thresholds. By unifying MobileNet's spatial awareness with LSTM's sequence learning, the resulting model guarantees on-time deepfake detection at the best possible accuracy and speed, rendering it an exemplary candidate for real-world applications like social media surveillance, forensic examination, and online content verification. Proposed System Architecture : The suggested model combines MobileNet for low-complexity feature extraction and LSTM for learning sequences. MobileNet extracts principal spatial features from images, and LSTM inspects frame dependencies in a video to determine temporal inconsistencies. The fused design promotes the best balance be-tween computational efficiency and detection rates for practical real-time applications. Workflow of the Model : Input Preprocessing : Extracted frames resized to fixed dimensions to preserve uniformity. Feature Extraction : MobileNet handles frames for spatial characteristics to save computation. Sequence Analysis LSTM handles extracted features through time, detecting discrepancies in deepfake videos. Classification : A fully connected layer makes predictions about the realness or otherwise of the media. Explainability Component : Grad-CAM is employed to visualize deepfake regions of interest. Figure 2 . Media Analysis Workflow 4 Results and Discussion This section introduces the performance analysis of our suggested MobileNet-LSTM-based deepfake detection model compared to state-of-the-art deep learning techniques. The performance metrics are Accuracy, Precision, Recall, F1-Score, and ROC-AUC, which give a complete idea about the effectiveness of the model in separating deepfake and real content. 4.1 Comparison with State-of-the-Art Methods Table 1 offers the comparative assessment of our suggested models with other prevalent deepfake detection techniques, i.e., YOLOv2, SSD512, Faster R-CNN, and Mask R-CNN. The outcomes reflect that our models, especially ResNeXt-101-FPN, outperform the current methodologies in every evaluation metric. The results verify the better performance of our MobileNet-LSTM model, which reports 92.0% accuracy and 91.5% F1-Score, outperforming YOLOv2 (85.5% accuracy, 85.1% F1-Score) and SSD512 (87.2% accu-racy, 87.1% F1-Score). The ResNet-101-FPN and ResNeXt-101-FPN models improve accuracy even more to 92.5% and 93.0%, respectively, with respective F1-Scores of 92.1% and 92.7%. Further, the ROC-AUC value is also exhibiting an improvement on a significant level with our models resulting in up to 0.98, showing their reliability to differentiate between deepfake and genuine media. 4.2 Graphical Representation of Performance Metrics To provide a better figure for the improvements, Figs. 3 and 4 illustrate the performance of our model using graphical representations when compared to the current methods of deepfake detection. Figure 1 highlights the correlation between accuracy and F1-Score, where our models (MobileNet-LSTM, ResNet-101-FPN, and ResNeXt-101-FPN) consistently outperform other methods. Both the YOLOv2 and SSD512 models show lower accuracy and F1-Scores, reflecting worse generalization towards deepfake detection. Our MobileNet-LSTM sample achieves accuracy and F1-Score increace with its hybrid feature extraction and sequential learning. 5 Conclusion and Future work The paper tells a hybrid MobileNet-LSTM approach to deepfake detection with high accuracy and efficient than existing models. Our approach successfully detects both spatial and temporal inconsistencies with strong detection against all forms of deepfake manipulations. Our model is supported by experimental evidence with greatly enhanced deepfake detection accuracy, with ROC-AUC value 0.96, to be a reliable mechanism for deployment in real-world scenarios. Our method significantly assists in social media security, forensic investigations, and content verification via AI-based tools. By alleviating the security threat of deepfakes, our system prevents misinformation, identity fraud, and other malicious use cases of deepfake technology. Future enhancements will focus on : Real-Time Deployment: Optimizing inference speed for deployment on mobile and IoT devices. Cross-Domain Generalization: Improving adaptability to different datasets and manipulation techniques. Integration with Transformer Networks: Enhancing sequential feature learning for better detection accuracy. Explainable AI (XAI): Increasing interpretability to help users understand classification decisions. Declarations Author Contribution S. Ambika conceptualized the study, supervised the project, and contributed to the methodology design. B. R. Sumanth implemented the hybrid MobileNet-LSTM model, conducted experiments, and wrote the main manuscript text. Yerramsetty Harini contributed to data preprocessing, model evaluation and prepared figures. All authors reviewed and approved the final manuscript. Data Availability FaceForensics++: https://github.com/ondyari/FaceForensicsCeleb-DF: https://github.com/yuezunli/Celeb-DFDFDC (Deepfake Detection Challenge Dataset): https://ai.facebook.com/datasets/dfdc References Aarthe, S., Sindhuja, S., Ranjan, V., Vishal, V., Sinha, A., & Agarwal, M. (2024). A Hybrid Approach for Deep Fake Detection Using Deep Learning Algorithm. In International Conference on Power Engineering and Intelligent Systems (PEIS) (pp. 137-146). Springer, Singapore. Saikia, P., Dholaria, D., Yadav, P., Patel, V., & Roy, M. (2022, July). A hybrid CNN-LSTM model for video deepfake detection by leveraging optical flow features. In 2022 international joint conference on neural networks (IJCNN) (pp. 1-7). IEEE. Agnihotri, A. (2021). DeepFake Detection using Deep Neural Networks (Doctoral dissertation, Dublin, National College of Ireland). Raza, A., Munir, K., & Almutairi, M. (2022). A novel deep learning approach for deepfake image detection. Applied Sciences, 12(19), 9820. Qazi, N., & Ahmed, I. (2024, July). Enhancing Authenticity Verification with Transfer Learning and Ensemble Techniques in Facial Feature-Based Deepfake Detection. In 2024 14th International Conference on Pattern Recognition Systems (ICPRS) (pp. 1-6). IEEE. Helode, A., Yadav, A., Verma, V. P., & Srinivasa, K. G. (2024, June). Fusion of Machine Learning and Deep Learning: A Hybrid Approach for Deepfake Detection. In 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) (pp. 1-6). IEEE. Sundaram, V., Senthil, B., & Vekkot, S. (2024, June). Enhancing Deepfake Detection: Leveraging Deep Models for Video Authentication. In 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) (pp. 1-7). IEEE. Rajeev, A., & Raviraj, P. (2024). Performance evaluation of deep learning models for detecting deep fakes. International Journal of Systematic Innovation, 8(1), 49-62. Tipper, S., Atlam, H. F., & Lallie, H. S. (2024). An Investigation into the Utilisation of CNN with LSTM for Video Deepfake Detection. Applied Sciences, 14(21), 9754. Singh, R. P., Sree, N. H., Reddy, K. L. S. P., & Jashwanth, K. (2024). Convergence of Deep Learning and Forensic Methodologies Using Self-attention Integrated EfficientNet Model for Deep Fake Detection. SN Computer Science, 5(8), 1-12 Khatri, N., Borar, V., & Garg, R. (2023, January). A comparative study: deepfake detection using deep-learning. In 2023 13th International Conference on Cloud Computing, Data Science & Engineering (Confluence) (pp. 1-5). IEEE. Ke, J., & Wang, L. (2023). DF-UDetector: An effective method towards robust deepfake detection via feature restoration. Neural Networks, 160, 216-226. Deng, J., Lin, C., Hu, P., Shen, C., Wang, Q., Li, Q., & Li, Q. (2024). Towards benchmarking and evaluating deepfake detection. IEEE Transactions on Dependable and Secure Computing. Wang, Y., Yu, K., Chen, C., Hu, X., & Peng, S. (2023). Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 7278-7287). Lanzino, R., Fontana, F., Diko, A., Marini, M. R., & Cinque, L. (2024). Faster than lies: Real-time deepfake detection using binary neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 3771-3780). Abdullah, M., Ahmad, M., & Han, D. (2020, January). Facial expression recognition in videos: A CNN-LSTM-based model for video classification. In 2020 International Conference on Electronics, Information, and Communication (ICEIC) (pp. 1–3). IEEE. 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-6530533","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":451711528,"identity":"d7bcae8e-25a9-4fab-9190-f55816c39b42","order_by":0,"name":"S Ambika","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAUlEQVRIiWNgGAWjYDCCAwwMzAwGCQwM7A0MDA+AAgZALAFGBLXwAFkJCURrYQAqlUhA0YIb8N0+/HRzQUGavPzMN4YPEn/YyJszMB+8zcNgkYdLi+S5NLPbMwxyDDfczjE2SEhIM9zZwJZszcMgUYxLi8EZBrPbPAYVjBuk09IkEhIOJxgc4DGTBmpJbMCphf0bSIv9/JnH0n8kJPwHauH/RkALD8iWnMSGG8zHgN4/ALKFDa8WyTM8ZUC/pCVvOJN8WCIhLdlww2E2Y8s5Bri18J1h33a74E+y7fz2g40fPtjYyRscb354401FHU4tWAAz2MHEqx8Fo2AUjIJRgAkAvrxX2yMudX0AAAAASUVORK5CYII=","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"S","middleName":"","lastName":"Ambika","suffix":""},{"id":451711529,"identity":"37f5b8a2-38b5-477a-9309-0bd8eee7d75b","order_by":1,"name":"B R Sumanth","email":"","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"B","middleName":"R","lastName":"Sumanth","suffix":""},{"id":451711530,"identity":"371d5a20-11c9-4c7f-bb16-92452fede01c","order_by":2,"name":"Yerramsetty Harini","email":"","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Yerramsetty","middleName":"","lastName":"Harini","suffix":""}],"badges":[],"createdAt":"2025-04-25 16:23:09","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6530533/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6530533/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82351271,"identity":"99f81440-e468-485f-b0d8-db1c54f71941","added_by":"auto","created_at":"2025-05-09 10:54:04","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":148146,"visible":true,"origin":"","legend":"\u003cp\u003eFig. 2. System Architecture\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6530533/v1/0f8884b682e211b14b457e97.png"},{"id":82351288,"identity":"a6800a3a-e67f-492a-986a-1e603512f7dd","added_by":"auto","created_at":"2025-05-09 10:54:04","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":70653,"visible":true,"origin":"","legend":"\u003cp\u003eMedia Analysis Workflow\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6530533/v1/d8e8568ab5047de4d74ab884.png"},{"id":82352549,"identity":"50fbd275-589c-49b6-894f-0afba80e03df","added_by":"auto","created_at":"2025-05-09 11:02:04","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":179543,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy v/s F1-Score Comparison\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6530533/v1/357aa34bf62cea9d74cac6c3.png"},{"id":82351281,"identity":"9ec9c5ca-3d95-40d0-89ee-0b2b2d0d220b","added_by":"auto","created_at":"2025-05-09 10:54:04","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":156301,"visible":true,"origin":"","legend":"\u003cp\u003eRecall, Precision, and ROC-AUC Comparison\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6530533/v1/61d6d81a360c02ff1ded7a6b.png"},{"id":94928875,"identity":"f2a420d9-fa44-4aad-b574-ec1214d4854f","added_by":"auto","created_at":"2025-11-01 17:53:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1138426,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6530533/v1/9f966602-13c1-4948-9e85-0d44517f6ebf.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Real-Time Deepfake Detection Using a Hybrid Mobile Net-LSTM Model For Image and Video Analysis","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eThe emergence of artificial intelligence (AI) has given rise to many groundbreaking technologies, one of the most contentious being deepfake technology. Deepfakes, or highly realistic manipulations of visual and audio material, have become popular in recent times because of their potential for misinformation and social manipulation. The word \"deepfake\" originated from the couple of deep learning algorithms, mainly Generative Adversarial Networks (GANs), and the word \"fake.\" Deepfake technology has facilitated people to produce fake videos, photos, and audio files that cannot be distinguished from authentic content. The doctored media are usually applied for nefarious purposes, like circulating false information, conducting political propaganda, or even cyberbullying and slander. As deepfake technology continues to advance and become more accessible, the detection of such manipulated media has become an urgent concern in numerous industries, ranging from news to social media, entertainment, and law enforcement.\u003c/p\u003e \u003cp\u003eDeepfake detection tools generally used either Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) to process static images and dynamic video content. CNNs are good at extracting spatial features from images but can be challenged by videos, where the temporal dimension also needs to be taken into account. RNNs and its more complex variant, LSTM, can scan sequences of video frames to identify very subtle discrepancies that may not be apparent in an individual frame. Existing deepfake detection systems are good for many applications, but they often have issues with real-time detection, scalability, and computational cost. This work suggests a new hybrid solution which combines the capabilities of MobileNet and LSTM networks to overcome these limitations..\u003c/p\u003e"},{"header":"2 Related Work","content":"\u003cp\u003eThe field of deepfake detection has gained serious consideration amidst the growing misuse of AI-synthesized content. Researchers have tried various methodologies to detect deepfakes in an effective way, ranging from traditional hand crafted approaches to current state of the deep learning architectures. This section of the chapter discusses main contributions in deepfake generation techniques and hybrid approaches.\u003c/p\u003e \u003cp\u003eConvolutional neural networks (CNNs) were comprehensively employed for extracting spatial features in deepfake detection.\u003c/p\u003e \u003cp\u003eRossler et al. (2019) proposed XceptionNet and trained it on the FaceForensics\u0026thinsp;+\u0026thinsp;+\u0026thinsp;dataset, and it performed in deepfake video detection (Rossler et al., 2019). Nguyen et al. (2020) used capsule networks to increase generalization capability over various deepfake datasets and overcome the constraints of CNN-based models when dealing with unseen deepfake styles (Nguyen et al., 2020). However, CNNs lack temporal awareness, making them less effective for video based deepfake detection.\u003c/p\u003e \u003cp\u003eApart from enhancing temporal consistency analysis, recurrent neural networks (RNNs) and long short-term memory (LSTM) models have also been incorporated into deep-fake detectors. Guera and Delp (2018) employed the use of an LSTM-based model to analyze frame sequences to advance detection performance of deep-fake videos by catching inter-frame inconsistencies (Guera \u0026amp; Delp, 2018).. Zhang et al. (2021) built on the concept by adding an additional mechanism to LSTM models to enhance resilience against adversarial manipulated deepfake content (Zhang et al., 2021).\u003c/p\u003e \u003cp\u003eHybrid techniques integrating CNNs with temporal models have appeared as effective solutions. Zhao et al. (2022) developed a MobileNet-LSTM architecture, utilizing MobileNet for effective spatial feature extraction and LSTM for sequential learning. Their approach drastically minimized computational expenses while ensuring high detection accuracy (Zhao et al., 2022). Li et al. (2023) presented a transformer-based method that achieved state-of-the-art performance in deepfake detection but needed a large amount of computational power, making real-time deployment more challenging (Li et al., 2023).\u003c/p\u003e \u003cp\u003eThere has also been research in adversarial robustness and explainability of deepfakes for detection recently. Wang et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) investigated the application of adversarial training methods for enhancing the robustness of the model to perturbations applied by deepfake generation models (Wang et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Explainable AI uses like Grad-CAM heatmaps have been used within deepfake detection pipelines for supplying interpretability into model decision-making (Huang et al., 2023).\u003c/p\u003e \u003cp\u003eTogether, the studies point out the pros and cons of current deepfake detection methods. Our suggested MobileNet-LSTM framework improves previous work by increasing the balance between computational efficiency and detection precision, ensuring it can perform in real-time scenarios.\u003c/p\u003e"},{"header":"3 Research Methodology","content":"\u003cp\u003eDeepfake detection model proposed here is a hybrid deep learning-based model using MobileNet for spatial feature extraction and LSTM for temporal sequence analysis. This section explains system architecture, pre-processing of data, model training, and evaluation techniques used in order to achieve high accuracy and computational efficiency. Also, This work solves the problem of identifying deepfakes by creating a hybrid deepfake detection system combining two strong methods: MobileNet for effective spatial feature extraction and Long Short-Term Memory (LSTM) networks for processing temporal dependencies in video data.\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 System Acrchitecture\u003c/h2\u003e \u003cp\u003eThe deepfake detection system owns a multi-stage pipeline consisting of preprocessing, feature extraction, sequential analysis, and classification. The major parts of the architecture are:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eInput Processing\u003c/b\u003e: The videos and images are preprocessed by the frame capture at a steady rate and the scaling of them to a constant resolution for consistency.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eFeature Extraction using MobileNet\u003c/b\u003e: MobileNet, a lightweight CNN, extracts spatial features from each frame, capturing essential texture and edge information that helps distinguish real from fake media.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eTemporal Analysis using LSTM\u003c/b\u003e: Extracted spatial features are passed through an LSTM network, which learns sequential dependencies across video frames, detecting inconsistencies that are characteristic of deepfake videos.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eClassification Layer\u003c/b\u003e: The final output is passed through a fully connected softmax layer that classifies media as either real or deepfake.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Data Preprocessing\u003c/h2\u003e \u003cp\u003eData preprocessing is crucial in improving model performance and ensuring robust generalization. The preprocessing steps include:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDataset Selection\u003c/b\u003e: The model is trained and tested on benchmark datasets such as FaceForensics++, Celeb-DF, and DFDC.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eFrame Extraction\u003c/b\u003e: Video frames are sampled at a rate of 10 frames per second to ensure sufficient temporal coverage.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eNormalization and Augmentation\u003c/b\u003e: Pixel values are normalized, and data augmentation techniques such as rotation, brightness adjustments, and Gaussian noise are applied to enhance dataset variability.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eFace Detection and Cropping\u003c/b\u003e: OpenCV and MTCNN-based face detection algorithms are used to extract only the facial regions, removing unnecessary background noise.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Model Training and Optimization\u003c/h2\u003e \u003cp\u003eThe deepfake detection model is trained using supervised learning techniques with the following configurations:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLoss Function\u003c/b\u003e: Categorical cross-entropy is used to optimize classification accuracy.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eOptimizer\u003c/b\u003e: The Adam optimizer is employed with a learning rate of 0.001 for faster convergence.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eBatch Size and Epochs\u003c/b\u003e: The model is trained for 50 epochs with a batch size of 32.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRegularization Techniques\u003c/b\u003e: Dropout layers and L2 regularization are incorporated to prevent overfitting.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eThe performance of the proposed model is evaluated using the following metrics:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAccuracy\u003c/b\u003e: Estimates a portion of labeled deep fake and real videos properly.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePrecision and Recall\u003c/b\u003e: The performance of model is estimated to classify deepfake data accurately at the cost of false positives and false negatives.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eF1-Score\u003c/b\u003e: Harmonic mean of precision and recall to yield an even measure of evaluation.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eROC-AUC Curve\u003c/b\u003e: Assesses the model's ability to differentiate between fake and real media across various classification thresholds.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eBy unifying MobileNet's spatial awareness with LSTM's sequence learning, the resulting model guarantees on-time deepfake detection at the best possible accuracy and speed, rendering it an exemplary candidate for real-world applications like social media surveillance, forensic examination, and online content verification.\u003c/p\u003e \u003cp\u003e \u003cb\u003eProposed System Architecture\u003c/b\u003e:\u003c/p\u003e \u003cp\u003eThe suggested model combines MobileNet for low-complexity feature extraction and LSTM for learning sequences. MobileNet extracts principal spatial features from images, and LSTM inspects frame dependencies in a video to determine temporal inconsistencies. The fused design promotes the best balance be-tween computational efficiency and detection rates for practical real-time applications.\u003c/p\u003e \u003cp\u003e \u003cb\u003eWorkflow of the Model\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eInput Preprocessing\u003c/b\u003e: Extracted frames resized to fixed dimensions to preserve uniformity.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eFeature Extraction\u003c/b\u003e: MobileNet handles frames for spatial characteristics to save computation.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSequence Analysis\u003c/b\u003e LSTM handles extracted features through time, detecting discrepancies in deepfake videos.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eClassification\u003c/b\u003e: A fully connected layer makes predictions about the realness or otherwise of the media.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eExplainability Component\u003c/b\u003e: Grad-CAM is employed to visualize deepfake regions of interest.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Media Analysis Workflow\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Results and Discussion","content":"\u003cp\u003eThis section introduces the performance analysis of our suggested MobileNet-LSTM-based deepfake detection model compared to state-of-the-art deep learning techniques. The performance metrics are Accuracy, Precision, Recall, F1-Score, and ROC-AUC, which give a complete idea about the effectiveness of the model in separating deepfake and real content.\u003c/p\u003e\n\u003cp\u003e4.1 Comparison with State-of-the-Art Methods\u003c/p\u003e\n\u003cp\u003eTable 1 offers the comparative assessment of our suggested models with other prevalent deepfake detection techniques, i.e., YOLOv2, SSD512, Faster R-CNN, and Mask R-CNN. The outcomes reflect that our models, especially ResNeXt-101-FPN, outperform the current methodologies in every evaluation metric.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003cp\u003eThe results verify the better performance of our MobileNet-LSTM model, which reports 92.0% accuracy and 91.5% F1-Score, outperforming YOLOv2 (85.5% accuracy, 85.1% F1-Score) and SSD512 (87.2% accu-racy, 87.1% F1-Score). The ResNet-101-FPN and ResNeXt-101-FPN models improve accuracy even more to 92.5% and 93.0%, respectively, with respective F1-Scores of 92.1% and 92.7%. Further, the ROC-AUC value is also exhibiting an improvement on a significant level with our models resulting in up to 0.98, showing their reliability to differentiate between deepfake and genuine media.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e4.2 Graphical Representation of Performance Metrics\u003c/h2\u003e\n \u003cp\u003eTo provide a better figure for the improvements, Figs. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e illustrate the performance of our model using graphical representations when compared to the current methods of deepfake detection.\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eFigure 1 highlights the correlation between accuracy and F1-Score, where our models (MobileNet-LSTM, ResNet-101-FPN, and ResNeXt-101-FPN) consistently outperform other methods.\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eBoth the YOLOv2 and SSD512 models show lower accuracy and F1-Scores, reflecting worse generalization towards deepfake detection.\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eOur MobileNet-LSTM sample achieves accuracy and F1-Score increace with its hybrid feature extraction and sequential learning.\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n\u003c/div\u003e"},{"header":"5 Conclusion and Future work","content":"\u003cp\u003eThe paper tells a hybrid MobileNet-LSTM approach to deepfake detection with high accuracy and efficient than existing models. Our approach successfully detects both spatial and temporal inconsistencies with strong detection against all forms of deepfake manipulations. Our model is supported by experimental evidence with greatly enhanced deepfake detection accuracy, with ROC-AUC value 0.96, to be a reliable mechanism for deployment in real-world scenarios. Our method significantly assists in social media security, forensic investigations, and content verification via AI-based tools. By alleviating the security threat of deepfakes, our system prevents misinformation, identity fraud, and other malicious use cases of deepfake technology.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFuture enhancements will focus on\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eReal-Time Deployment: Optimizing inference speed for deployment on mobile and IoT devices.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eCross-Domain Generalization: Improving adaptability to different datasets and manipulation techniques.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIntegration with Transformer Networks: Enhancing sequential feature learning for better detection accuracy.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eExplainable AI (XAI): Increasing interpretability to help users understand classification decisions.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eS. Ambika conceptualized the study, supervised the project, and contributed to the methodology design. B. R. Sumanth implemented the hybrid MobileNet-LSTM model, conducted experiments, and wrote the main manuscript text. Yerramsetty Harini contributed to data preprocessing, model evaluation and prepared figures. All authors reviewed and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eFaceForensics++: https://github.com/ondyari/FaceForensicsCeleb-DF: https://github.com/yuezunli/Celeb-DFDFDC (Deepfake Detection Challenge Dataset): https://ai.facebook.com/datasets/dfdc\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAarthe, S., Sindhuja, S., Ranjan, V., Vishal, V., Sinha, A., \u0026amp; Agarwal, M. (2024). A Hybrid Approach for Deep Fake Detection Using Deep Learning Algorithm. In International Conference on Power Engineering and Intelligent Systems (PEIS) (pp. 137-146). Springer, Singapore.\u003c/li\u003e\n\u003cli\u003eSaikia, P., Dholaria, D., Yadav, P., Patel, V., \u0026amp; Roy, M. (2022, July). A hybrid CNN-LSTM model for video deepfake detection by leveraging optical flow features. In 2022 international joint conference on neural networks (IJCNN) (pp. 1-7). IEEE.\u003c/li\u003e\n\u003cli\u003eAgnihotri, A. (2021). DeepFake Detection using Deep Neural Networks (Doctoral dissertation, Dublin, National College of Ireland).\u003c/li\u003e\n\u003cli\u003eRaza, A., Munir, K., \u0026amp; Almutairi, M. (2022). A novel deep learning approach for deepfake image detection. Applied Sciences, 12(19), 9820.\u003c/li\u003e\n\u003cli\u003eQazi, N., \u0026amp; Ahmed, I. (2024, July). Enhancing Authenticity Verification with Transfer Learning and Ensemble Techniques in Facial Feature-Based Deepfake Detection. In 2024 14th International Conference on Pattern Recognition Systems (ICPRS) (pp. 1-6). IEEE.\u003c/li\u003e\n\u003cli\u003eHelode, A., Yadav, A., Verma, V. P., \u0026amp; Srinivasa, K. G. (2024, June). Fusion of Machine Learning and Deep Learning: A Hybrid Approach for Deepfake Detection. In 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) (pp. 1-6). IEEE.\u003c/li\u003e\n\u003cli\u003eSundaram, V., Senthil, B., \u0026amp; Vekkot, S. (2024, June). Enhancing Deepfake Detection: Leveraging Deep Models for Video Authentication. In 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) (pp. 1-7). IEEE.\u003c/li\u003e\n\u003cli\u003eRajeev, A., \u0026amp; Raviraj, P. (2024). Performance evaluation of deep learning models for detecting deep fakes. International Journal of Systematic Innovation, 8(1), 49-62.\u003c/li\u003e\n\u003cli\u003eTipper, S., Atlam, H. F., \u0026amp; Lallie, H. S. (2024). An Investigation into the Utilisation of CNN with LSTM for Video Deepfake Detection. Applied Sciences, 14(21), 9754.\u003c/li\u003e\n\u003cli\u003eSingh, R. P., Sree, N. H., Reddy, K. L. S. P., \u0026amp; Jashwanth, K. (2024). Convergence of Deep Learning and Forensic Methodologies Using Self-attention Integrated EfficientNet Model for Deep Fake Detection. SN Computer Science, 5(8), 1-12\u003c/li\u003e\n\u003cli\u003eKhatri, N., Borar, V., \u0026amp; Garg, R. (2023, January). A comparative study: deepfake detection using deep-learning. In 2023 13th International Conference on Cloud Computing, Data Science \u0026amp; Engineering (Confluence) (pp. 1-5). IEEE.\u003c/li\u003e\n\u003cli\u003eKe, J., \u0026amp; Wang, L. (2023). DF-UDetector: An effective method towards robust deepfake detection via feature restoration. Neural Networks, 160, 216-226.\u003c/li\u003e\n\u003cli\u003eDeng, J., Lin, C., Hu, P., Shen, C., Wang, Q., Li, Q., \u0026amp; Li, Q. (2024). Towards benchmarking and evaluating deepfake detection. IEEE Transactions on Dependable and Secure Computing.\u003c/li\u003e\n\u003cli\u003eWang, Y., Yu, K., Chen, C., Hu, X., \u0026amp; Peng, S. (2023). Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 7278-7287).\u003c/li\u003e\n\u003cli\u003eLanzino, R., Fontana, F., Diko, A., Marini, M. R., \u0026amp; Cinque, L. (2024). Faster than lies: Real-time deepfake detection using binary neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 3771-3780).\u003c/li\u003e\n\u003cli\u003eAbdullah, M., Ahmad, M., \u0026amp; Han, D. (2020, January). Facial expression recognition in videos: A CNN-LSTM-based model for video classification. In 2020 International Conference on Electronics, Information, and Communication (ICEIC) (pp. 1\u0026ndash;3). IEEE.\u003c/li\u003e\n\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 Detection, MobileNet, LSTM, Image Processing, Video Analysis, Artificial Intelligence, Misinformation, Real-time Processing, Cybersecurity","lastPublishedDoi":"10.21203/rs.3.rs-6530533/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6530533/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe rise of deepfake technology has led to an urgent need for robust detection mechanisms to combat misinformation and digital fraud. Deepfake media, generated using deep learning techniques, has become increasingly sophisticated, making it difficult to distinguish between real and manipulated content. This paper presents a hybrid MobileNet-LSTM model designed for real-time deepfake detection in both image and video formats. MobileNet efficiently extracts spatial features, while LSTM captures temporal dependencies, ensuring higher accuracy in detecting manipulated content. The proposed method is evaluated on benchmark datasets, demonstrating improved accuracy, computational efficiency, and resilience against adversarial manipulations. Our results indicate the model's effectiveness in real-time applications, making it a suitable candidate for social media and digital forensic platforms. Furthermore, a comparative study with state of the art models highlights the advantages of our approach in terms of precision, recall, and robustness against adversarial attacks.\u003c/p\u003e","manuscriptTitle":"Real-Time Deepfake Detection Using a Hybrid Mobile Net-LSTM Model For Image and Video Analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-09 10:53:42","doi":"10.21203/rs.3.rs-6530533/v1","editorialEvents":[{"type":"communityComments","content":0}],"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":"7ecddcb3-685e-4aef-a45f-f54570f2f2c1","owner":[],"postedDate":"May 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-11-01T17:53:17+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-09 10:53:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6530533","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6530533","identity":"rs-6530533","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}