Secure Hardware Assurance Using Visual AI on AOI Imaging of Electronic Assemblies

Secure Hardware Assurance Using Visual AI on AOI Imaging of Electronic Assemblies | 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 Secure Hardware Assurance Using Visual AI on AOI Imaging of Electronic Assemblies Eyal Weiss This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7197469/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 09 Mar, 2026 Read the published version in Journal of Hardware and Systems Security → Version 1 posted 10 You are reading this latest preprint version Abstract Ensuring the authenticity and integrity of electronic assemblies is increasingly critical as hardware-based attacks and unauthorized component modifications become more sophisticated. Conventional inspection systems, whether rule-based AOI or traceability logs, offer limited protection against subtle or intentional tampering. This paper introduces a deep learning–based framework for secure hardware assurance that operates directly on AOI image data, enabling autonomous, full-coverage verification of every component on the board. The method is built on two previously patented systems: one for component authentication via visual fingerprinting—independent of top marking—and another for contextual decoding of top marking codes. These systems have been deployed across tens of SMT lines, generating over 5 billion production-grade inspections. By integrating and extending these capabilities, the system performs bottom-up part analysis and top-down layout validation, identifying substitutions, rework, and tampering, without requiring electrical probing, golden boards, or metadata. Results show > 99% detection accuracy and sub-second inspection times, enabling secure, image-only verification suitable for in-line or forensic use. This work bridges the gap between software protections and physical hardware trust, transforming AOI images into a practical security and compliance enforcement tool. visual inspection anomaly detection secure hardware assurance deep learning automated optical inspection (AOI) component authentication PCB assem-bly counterfeit detection manufacturing integrity hardware cybersecurity Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction The reliability and security of modern electronic systems are increasingly challenged by component-level issues, including counterfeit parts, unauthorized modifications, and assembly inconsistencies [ 1 – 3 ]. These risks are particularly critical in domains such as aerospace, medical, energy infrastructure, and defense electronics, where undetected deviations can result in severe functional failures or national security breaches [ 4 , 5 ]. Although Automated Optical Inspection (AOI) systems are widely used in electronics manufacturing, their scope remains limited to rule-based validation of basic attributes, component presence, orientation, polarity, leaving more subtle or undocumented modifications undetected [ 6 – 13 ]. Meanwhile, high-resolution AOI images, captured routinely during production, contain rich visual information that could reveal anomalies in shape, markings, package geometry, lead configuration, and solder interfaces. These features are rarely interpreted beyond pass/fail rules and are completely disconnected from security assurance or supply chain verification efforts [ 6 , 10 ]. This limitation has come into focus in recent years. A May 2025 investigative report by Reuters revealed the presence of rogue wireless modules embedded in Chinese-manufactured solar inverters deployed across U.S. infrastructure [ 14 ]. These undocumented components were invisible to conventional inspection and enabled covert remote access that bypassed all software protections. A similar case, the widely discussed “Big Hack”, involved allegations of tiny unauthorized devices implanted onto server boards to facilitate cyber-espionage [ 1 ]. Such incidents underscore the growing disconnect between digital cybersecurity protections and the physical verification of hardware. Current approaches to [ 4 , 15 , 16 ] trust rely heavily on documentation, sample testing, and presumed supplier integrity, rather than verifying the actual physical materials present on the board [ 3 ]. This results in a security blind spot: if a component is visually wrong, but electrically equivalent, or placed in an uninspected area, it is likely to go undetected. Rule-based inspection systems and traceability frameworks were never designed to catch these scenarios. This paper addresses that gap by introducing a practical, image-based method to perform component-level [ 17 ] using the same AOI data already generated during production. Our approach enables scalable, non-invasive detection of unauthorized component [ 2 , 16 , 18 – 21 ]changes and structural anomalies, providing a critical layer of trust at the physical level, without requiring [ 18 ], golden boards, or complete traceability metadata [ 22 – 24 ]. This work builds on two foundational capabilities previously developed and patented by our team. The first involves visual AI–based authentication of electronic components using their physical fingerprint, geometry, lead structure, and surface finish, to identify the manufacturer regardless of top markings [ 25 , 26 ]. The second enables contextual decoding of top marking codes to extract part numbers, lot/date codes, and traceability metadata. These methods have been field-proven through deployment on tens of SMT production lines, accumulating over 5 billion component inspections [ 27 ]. The current framework integrates and extends these capabilities toward full-board secure hardware assurance. 2. Method The proposed inspection framework is designed to analyze high-resolution images generated by Automated Optical Inspection (AOI) systems [ 10 ], enabling component-level verification and anomaly detection through deep learning. It integrates object detection for identifying and localizing all components on a board, semantic segmentation [ 28 ]for analyzing component geometry and solder interfaces, and a downstream anomaly classifier that evaluates deviations from expected patterns. This architecture allows the system to detect visual anomalies that indicate tampering, substitution, or undocumented rework, with no reliance on electrical measurements, rule-based tolerances, or golden board comparisons. The workflow, outlined in Fig. 1 , includes dataset acquisition, model training, and real-time inspection on new boards. The framework operates along two parallel tracks: A bottom-up component analysis, where each part is visually fingerprinted based on body shape, leads or balls, logo, top marking, and package features [ 25 , 26 ]; A top-down layout inspection, where the entire board is evaluated for missing, added, or misplaced components relative to learned spatial patterns. This dual-layer strategy enables the system to detect both local and global deviations with high sensitivity and flexibility across board types. 2.1. Dataset and Training Training data was acquired from two sources: Clean production boards, imaged inline during assembly with consistent lighting and camera setup. Degraded boards, collected from scrap lots and recycling centers, imaged in both cleaned and uncleaned conditions to capture real-world noise: corrosion, oxidation, broken components, dirt, and surface wear. The use of multi-domain data ensures generalization across typical production variability and post-deployment board conditions. The dataset was expanded using a bootstrapped training strategy [ 29 ]: Initial models trained on pristine images were used to infer component locations and masks on degraded boards. These predictions were manually reviewed and corrected to generate new labeled examples. The updated data was used to retrain the models. This cycle was repeated until detection and segmentation stabilized across pristine and degraded conditions. This approach eliminates the need for golden boards or exhaustive pre-labeling and enables the model to handle image noise, occlusion, and structural damage that typically degrade rule-based inspection systems. The dataset and model training strategy build on two prior systems developed by our group: (1) a bottom-view authentication engine used to identify component source based on physical characteristics, and (2) a top-view marking decoder trained to read and interpret manufacturer-specific markings. Both systems have been deployed at scale in production and provided the initial models and annotation tools used in this work. This multi-domain dataset supports model generalization across pristine, degraded, and field-recovered board conditions, critical for real-world deployment where production variability and contamination are common. 2.2. Object Detection Component-level object detection is performed using the YOLOv12 architecture [ 28 , 30 ]. This model was chosen for its strong performance in dense scenes and its speed, enabling inference under 1000 ms per board. The detector is trained to localize and classify components by package type using CAD-aligned bounding boxes. Data augmentation techniques, flipping, brightness shifts, mild rotation, and Gaussian blu, were applied to improve generalization across lighting and contamination conditions. In deployment, the model outputs bounding boxes, class labels, and confidence scores for each component. These detections are used both for component fingerprinting and for guiding segmentation and anomaly classification. 2.3. Component Segmentation A semantic segmentation model complements detection by producing pixel-level masks for each component body and its solder pad regions. These masks enable analysis of footprint alignment, pad coverage, and geometric consistency with reference designs. As with detection, segmentation was trained using both pristine and degraded boards, bootstrapped through iterative retraining. This ensures that the model performs reliably even under noise and partial occlusion. The output masks serve two purposes: Geometry extraction for anomaly classification Layout validation against expected pad/footprint shapes This module enhances the system’s ability to detect subtle anomalies such as footprint mismatch, rework, or soldering defects. Output from the semantic segmentation model showing detailed masking of each component’s body (shown in blue) and the associated solder pad regions (highlighted in red). These masks enable fine-grained geometric comparisons with the golden reference. 2.4. Anomaly Classification An anomaly classifier receives features from the detection and segmentation stages. Each component is represented by a feature vector describing: Bounding box size and aspect ratio Pad and body mask shape metrics Relative position and spacing Texture descriptors and marking alignment These features are processed by a shallow CNN that assigns each part to one of several anomaly classes: missing, added, rotated, piggybacked, contaminated, unknown, or reworked. The classifier operates independently of detection confidence, allowing detection and classification decisions to be decoupled. To reduce false positives, low-confidence detections and geometric outliers are filtered or flagged. Spatial thresholds are learned from reference layouts and adjusted for each board type. 2.5. Framework Overview The full system operates as a modular pipeline: High-resolution AOI images are processed by the object detector. Detected regions are passed to the segmentation module. Combined features are evaluated by the anomaly classifier. Output anomalies are categorized and scored for downstream action (alerting, review, traceability linking, etc.). The system operates at production speed and requires no modification to the AOI hardware or imaging process. Once trained, it performs inspection autonomously, providing board-level results in sub-second timeframes. 3. Experimental Results and Case Studies To evaluate the performance of the proposed AOI image–based inspection framework, we conducted experiments across diverse datasets representing both clean production conditions and degraded, real-world boards. The evaluation focused on component detection accuracy, segmentation quality, anomaly classification, and overall robustness under varying visual conditions. 3.1. Detection and Segmentation Performance The object detection model, based on YOLOv12, was trained using three dataset types: Pristine AOI images, aligned with CAD data; Cleaned scrap boards, scanned under controlled lighting; Uncleaned scrap boards, representing worst-case conditions including oxidation, broken parts, and contamination. Model training followed the bootstrapped strategy described in Section 2 . Training convergence over 360 epochs is shown in Fig. 4 , which illustrates loss stability and generalization across the full dataset. Model performance on unseen boards is summarized in Table I. The model achieved an average F1 score of 0.90 across board types, with precision and recall remaining high even on degraded inputs. Table I – Detection Performance Across Board Conditions. Board Condition F1 Score Precision Recall Pristine AOI Image 0.96 0.96 0.96 Clean Scrap Board 0.92 0.89 0.96 Dirty Scrap Board 0.82 0.78 0.89 Average 0.90 0.88 0.94 These results confirm the detector’s resilience. Although performance dips under severe degradation, precision and recall remain strong, enabling reliable component localization for downstream analysis. The segmentation model was evaluated on 2,371 component instances across 241 AOI images. Results are shown in Table II. High mAP50 scores demonstrate accurate mask alignment, particularly on solder pads. Table II – Segmentation Accuracy by Class. Class Precision Recall mAP50 Body 0.911 0.919 0.961 Pad 0.975 0.934 0.978 All 0.943 0.927 0.969 Figure 5 . Precision-Confidence Curve for Segmentation. The model maintains stable behavior across confidence thresholds, with average segmentation precision around 0.92. 3.2. Real-World Case Studies The framework was applied to production boards to evaluate its effectiveness in detecting undocumented deviations. These real-world examples demonstrate its ability to catch anomalies that would bypass conventional traceability and AOI systems. In one case, a microcontroller was flagged for inconsistent markings. Visual comparison revealed it was sourced from a different manufacturer, despite electrical equivalence. The system flagged the anomaly based on shape and logo inconsistencies. Another board contained a passive component logged as Panasonic, but identified visually as a Vishay part. Package geometry and finish mismatches led to its classification as a substitution. A separate example involved a part logged as Bourns but unrecognized by the model. The system classified it as unknown, and manual inspection confirmed it was a counterfeit or unqualified device. Figure 6 shows layout-level anomalies detected through top-down board inspection. Components added, missing, or misplaced relative to the learned layout are clearly flagged. The system also detected piggybacking—a technique where unauthorized components are soldered atop legitimate ones to alter function. Figure 7 shows detection of such tampering on a board, where geometric irregularities triggered anomaly classification. In another case, a wire had been manually soldered between two nodes post-production, likely to override circuit behavior or inject a side-channel function. Figure 8 shows this anomaly, along with an obscuring sticker, both caught by layout-based anomaly detection. Finally, Fig. 9 shows a component substitution where a crystal oscillator was visually replaced with an alternate vendor’s part. The substitution was detected via shape, pad layout, and silkscreen misalignment. 3.3. Scalability and Throughput Performance was benchmarked across full board scans at standard AOI resolutions. The system achieves: > 99.3% anomaly detection accuracy < 0.5% false positive rate < 1% false negative rate < 1000 ms total processing time per board Component-level inference runs at under 10 ms per instance. This enables full-coverage inspection without disrupting throughput or requiring manual review. The framework integrates directly into post-AOI workflows and can also be applied retrospectively for forensics or supply chain audits. 4. Discussion This work introduces a novel approach to hardware assurance: a visual AI system that performs full-coverage component-level verification using only AOI image data. Unlike conventional systems, which rely on rule-based logic, golden boards, or traceability records, our method directly analyzes the visual identity and spatial context of every component on the board, enabling the detection of anomalies that were previously undetectable at scale. The framework described here is a direct evolution of two earlier, production-grade systems that separately addressed component-level authentication and marking interpretation. By unifying these capabilities into a board-level inspection architecture and training on a much broader dataset, we demonstrate, for the first time, scalable, component-by-component hardware assurance directly from AOI imagery. The core novelty lies in the dual-layer inspection strategy: A bottom-up component fingerprinting process, where each part is analyzed individually for shape, lead configuration, markings, and logos; and A top-down board-level consistency check, which learns spatial relationships across multiple known-good boards and flags layout anomalies such as missing, swapped, or added components. Together, these layers allow the system to detect both local and global modifications, including unauthorized substitutions, rework, and tampering, without requiring any electrical testing or metadata. To our knowledge, no prior work has demonstrated this level of scalable, image-only hardware verification on production boards. The system’s generalization ability is another key contribution. Through a bootstrapped training pipeline, it learns from both pristine and degraded board conditions, extending its performance beyond lab settings to real-world factory floors, RMA returns, or field-recovered units. This enables the method to maintain high accuracy even in the presence of oxidation, contamination, or partial damage, scenarios that defeat traditional rule-based AOI and even most learning-based defect classifiers. Additionally, the use of standard AOI imagery makes the method immediately deployable. It introduces no production bottlenecks and requires no changes to inspection equipment. This enables manufacturers to repurpose their existing image streams, not only for quality assurance but also for supply chain verification, provenance validation, and hardware security enforcement. While performance is strong, several practical limitations remain. The system depends on sufficient imaging quality, which can vary across AOI platforms. It also cannot determine intent, e.g., whether an anomaly stems from malicious tampering, unapproved rework, or an upstream substitution error. Expanding the component reference database and improving automated anomaly attribution will be key for scaling across product lines and sectors. Nonetheless, this framework represents a significant evolution in how manufacturing data can be used. By combining object detection, semantic segmentation, and anomaly classification in a modular pipeline, and grounding inspection in visual pattern learning rather than rule definition, the system provides a new pathway toward trusted electronics at the physical layer. This methodology repositions AOI imagery from a passive QA artifact to a forensic-grade security tool, enabling industries to verify not only whether components were placed, but whether the right components, from the right sources, were used. 5. Conclusion This work presents a novel, scalable framework for secure hardware assurance based entirely on AOI image data and deep learning. By analyzing each component’s visual identity, including shape, lead structure, and markings, and comparing board-wide layouts across known-good samples, the system detects subtle anomalies such as substitutions, unauthorized additions, or tampering, without the need for electrical testing or traceability metadata. Unlike existing inspection or authentication methods, which are limited by rules, golden boards, or metadata availability, our approach performs autonomous, full-coverage verification using data already captured in standard manufacturing workflows. This transforms AOI from a passive quality assurance step into an active layer of physical trust enforcement. The method has been validated across a broad range of conditions, from pristine boards to visually degraded and field-returned samples, demonstrating high accuracy, generalization, and fast inference suitable for in-line deployment. Its ability to detect deviations that evade conventional AOI or software-based security makes it well-suited also for critical sectors such as aerospace, defense, and medical electronics. By bridging the gap between image-based inspection and hardware cybersecurity, this system offers a new model for defending the physical supply chain, where hardware authenticity is no longer assumed, but verified. This framework integrates two previously patented systems: one for identifying component origin via visual fingerprinting, and another for decoding manufacturer-specific top markings. Both have been deployed across major production lines and trained on over 5 billion AOI images. By combining these proven capabilities with board-level spatial reasoning and anomaly detection, the presented solution delivers a new standard for hardware trust at the physical layer, scalable, autonomous, and grounded in real-world manufacturing data. Declarations Ethical Approval not applicable Funding not applicable Availability of data and materials This study does not rely on publicly available datasets. All data generated or analyzed during this study are either proprietary or derived from internal operational systems and are not publicly available due to confidentiality agreements. Reasonable requests for access to anonymized data supporting the findings of this study may be considered by the corresponding author, subject to approval and compliance with relevant data protection regulations. Author Contribution Eyal Weiss conceived the research, developed the methodology, conducted the experiments, prepared the figures, and wrote the manuscript. The author reviewed and approved the final version. References D. Mehta, H. Lu, O. P. Paradis, M. A. MS, M. T. Rahman, Y. Iskander, P. Chawla, D. L. Woodard, M. Tehranipoor, and N. Asadizanjani, ACM Journal on Emerging Technologies in Computing Systems (JETC) 16 , 1 (2020). M. M. Tehranipoor, U. Guin, and D. Forte, in Counterfeit Integrated Circuits (Springer, 2015), pp. 15–36. S. Akter, K. Khalil, and M. Bayoumi, IEEe Access 11 , 77543 (2023). H. Representatives, (2017). M. Tehranipoor, N. Pundir, N. Vashistha, and F. Farahmandi, Hardware Security Primitives (Springer, 2023). R. Ajax, (n.d.). A. Hemmati, P. Raoufi, and A. M. Rahmani, Neural Comput Appl 36 , 11461 (2024). M. M. M. Islam, J. I. Emon, K. Y. Ng, A. Asadpour, M. M. R. Al Aziz, M. L. Baptista, and J.-M. Kim, in Artificial Intelligence for Smart Manufacturing and Industry X. 0 (Springer, 2025), pp. 9–36. V. Reshadat and R. A. J. W. Kapteijns, in 2021 International Conference on Data and Software Engineering (ICoDSE) (IEEE, 2021), pp. 1–5. L. Meyer, B. Josefsson, K. Vrotsou, C. Westin, and J. Lundberg, in 2021 Ieee/Aiaa 40th Digital Avionics Systems Conference (Dasc) (IEEE, 2021), pp. 1–8. X. Tong, Z. Yu, X. Tian, H. Ge, and X. Wang, Front Comput Sci 16 , 1 (2022). Y.-S. Ni, W.-L. Chen, Y. Liu, M.-H. Wu, and J.-I. Guo, Sensors 24 , 5737 (2024). I.-C. Chen, R.-C. Hwang, and H.-C. Huang, Processes 11 , 775 (2023). Sarah Mcfarlane, Reuters (2025). A. Spieske and H. Birkel, Comput Ind Eng 158 , 107452 (2021). S. P. Gayialis, E. P. Kechagias, G. A. Papadopoulos, and D. Masouras, Sustainability (Switzerland) 14 , (2022). M. T. Rahman, Q. Shi, S. Tajik, H. Shen, D. L. Woodard, M. Tehranipoor, and N. Asadizanjani, in 2018 IEEE 3rd International Verification and Security Workshop (IVSW) (2018), pp. 93–102. M. M. Tehranipoor, U. Guin, and D. Forte, in Counterfeit Integrated Circuits (Springer, 2015), pp. 95–107. M. M. Tehranipoor, U. Guin, and D. Forte, in Counterfeit Integrated Circuits (Springer, 2015), pp. 75–93. L. W. Kessler and T. Sharpe, Printed Circuit Design & Fab 27 , 64 (2010). a S. Report, T (2011). G. M. Razak, L. C. Hendry, and M. Stevenson, Production Planning and Control 34 , 1114 (2023). P. Rosales, (2021). R. Schuitemaker and X. Xu, in Procedia CIRP (Elsevier B.V., 2020), pp. 700–705. E. Weiss, IEEE Trans Compon Packaging Manuf Technol 1 (2024). E. Weiss, SMTA (2021). E. Weiss and Z. Efrat, P-603538-PC (2021). Y. Li, Y. Duan, L. Duan, W. Xiang, and Q. Wu, in Medical Image Understanding and Analysis , edited by M. H. Yap, C. Kendrick, A. Behera, T. Cootes, and R. Zwiggelaar (Springer Nature Switzerland, Cham, 2024), pp. 143–159. J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Daniel Guo, M. Gheshlaghi Azar, B. Piot, K. Kavukcuoglu, R. Munos, and M. Valko, Bootstrap Your Own Latent A New Approach to Self-Supervised Learning (n.d.). Y. Li, S. Li, H. Du, L. Chen, D. Zhang, and Y. Li, IEEE Access 8 , 227288 (2020). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 09 Mar, 2026 Read the published version in Journal of Hardware and Systems Security → Version 1 posted Editorial decision: Revision requested 05 Oct, 2025 Reviews received at journal 27 Aug, 2025 Reviewers agreed at journal 09 Aug, 2025 Reviews received at journal 07 Aug, 2025 Reviewers agreed at journal 04 Aug, 2025 Reviewers agreed at journal 04 Aug, 2025 Reviewers invited by journal 04 Aug, 2025 Editor assigned by journal 03 Aug, 2025 Submission checks completed at journal 31 Jul, 2025 First submitted to journal 23 Jul, 2025 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-7197469","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":495960837,"identity":"8e36d52b-0f17-4c51-bd52-173e6e4c5b22","order_by":0,"name":"Eyal Weiss","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAw0lEQVRIiWNgGAWjYHACxgMMDDYMDBJAJg8YEQGAWtJI13IYroUw4JdIfnCAse28vMHt5mMf3jDYyRDUIjkjzQCo5bbhhjvHkmfOYUgmbJHBjQSglm23Ewxu5Bgz8zAwE9ZifyP9A1DLOZiWeiJskcgB2XIApuUwYS0SZ94UHEj8l2w4E+gXxjkGxwlr4W9P3/jgwxk7eb7bzYcZ3lRU2xPUwiCQwMAARAwKB8DuJKwBaM0BCC3fQIzqUTAKRsEoGJEAAGkiPR2eKpeHAAAAAElFTkSuQmCC","orcid":"","institution":"Cybord.ai","correspondingAuthor":true,"prefix":"","firstName":"Eyal","middleName":"","lastName":"Weiss","suffix":""}],"badges":[],"createdAt":"2025-07-23 14:38:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7197469/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7197469/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s41635-026-00173-5","type":"published","date":"2026-03-09T15:59:09+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":88416206,"identity":"1070cd69-588f-497e-a4f4-fc754552c5b9","added_by":"auto","created_at":"2025-08-06 08:58:36","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":206643,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the Inspection Framework Pipeline. The workflow includes acquisition of clean and scrap board datasets, initial training of YOLOv12 and segmentation models, bootstrapped retraining on degraded boards, and final anomaly detection based on detection and segmentation outputs.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/2f91fbabecc6b0d6c61cd06e.png"},{"id":88414945,"identity":"adf60b3a-f158-4a9e-9488-3fe7f2630f03","added_by":"auto","created_at":"2025-08-06 08:50:36","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":385517,"visible":true,"origin":"","legend":"\u003cp\u003eComponent detection on a pristine board. The top image shows a high-resolution AOI image processed by the object detection model, with red bounding boxes accurately localizing all visible components. The bottom image displays extracted metadata from the same image, illustrating how detection results serve as the basis for downstream anomaly analysis.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/bf9e8e366c79c641d356f177.png"},{"id":88414942,"identity":"cba51f3e-6e4e-4385-aca2-76cf640dfbcd","added_by":"auto","created_at":"2025-08-06 08:50:36","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":405562,"visible":true,"origin":"","legend":"\u003cp\u003eComponent-Level Segmentation Results.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/04f845c4e5a56c6280922028.png"},{"id":88416207,"identity":"2fe8d231-464d-4cf8-95cb-e5091dee3c48","added_by":"auto","created_at":"2025-08-06 08:58:36","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":88724,"visible":true,"origin":"","legend":"\u003cp\u003eTraining Dynamics Across 360 Epochs. Loss convergence curves for object detection, showing classification, regression, and localization losses. Steady convergence across clean and degraded boards supports the model’s generalization ability.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/af7c43657f9208a257d3811d.png"},{"id":88417026,"identity":"543a043d-0d3c-45f1-bbad-391aaf4834b9","added_by":"auto","created_at":"2025-08-06 09:06:36","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":28387,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ePrecision-Confidence Curve for Segmentation.\u003c/em\u003e The model maintains stable behavior across confidence thresholds, with average segmentation precision around 0.92.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/32da3c2df07ed3c08c92ea74.png"},{"id":88416212,"identity":"204339a5-85aa-492a-b6c9-7ccbf44462ef","added_by":"auto","created_at":"2025-08-06 08:58:36","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":416006,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eDetection of layout-level anomalies.\u003c/em\u003e Yellow highlights indicate visual deviations from expected spatial configuration, even outside typical AOI coverage zones.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/61f4b11a15badcd5d7ed23e1.png"},{"id":88416215,"identity":"e78d1a27-5048-4600-a579-68e487ad0f44","added_by":"auto","created_at":"2025-08-06 08:58:36","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":445938,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eDetection of unauthorized and misplaced components.\u003c/em\u003e Top: scattered or missing components. Bottom: piggybacked device (U_8) flagged due to deviation in shape and placement.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/6926b967675a3d86da15f42a.png"},{"id":88417028,"identity":"a630c5a3-8094-477e-bb0e-30a2f804c6af","added_by":"auto","created_at":"2025-08-06 09:06:36","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":1083652,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eDetection of physical modifications.\u003c/em\u003e Left: wire bridge added manually. Right: obscuring label adhered to the board surface. Both fall outside AOI-defined zones but are detected by the layout consistency engine.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/57a9905ae6f3367e5abac479.png"},{"id":88414962,"identity":"9c70aefd-67d1-42b2-95fd-67bb42ce6701","added_by":"auto","created_at":"2025-08-06 08:50:36","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":324658,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eUnauthorized component substitution.\u003c/em\u003e Left: reference distribution across 50 known-good boards. Right: test board with mismatched oscillator, flagged by the system.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/df3ad722f1848433d89e0ff5.png"},{"id":104739491,"identity":"228edc10-2b40-4a45-b7f0-128c56b7d821","added_by":"auto","created_at":"2026-03-16 16:07:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4022008,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7197469/v1/aacb0b6d-e5f8-4429-9cd6-7a36344cb148.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Secure Hardware Assurance Using Visual AI on AOI Imaging of Electronic Assemblies","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe reliability and security of modern electronic systems are increasingly challenged by component-level issues, including counterfeit parts, unauthorized modifications, and assembly inconsistencies [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. These risks are particularly critical in domains such as aerospace, medical, energy infrastructure, and defense electronics, where undetected deviations can result in severe functional failures or national security breaches [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Although Automated Optical Inspection (AOI) systems are widely used in electronics manufacturing, their scope remains limited to rule-based validation of basic attributes, component presence, orientation, polarity, leaving more subtle or undocumented modifications undetected [\u003cspan additionalcitationids=\"CR7 CR8 CR9 CR10 CR11 CR12\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMeanwhile, high-resolution AOI images, captured routinely during production, contain rich visual information that could reveal anomalies in shape, markings, package geometry, lead configuration, and solder interfaces. These features are rarely interpreted beyond pass/fail rules and are completely disconnected from security assurance or supply chain verification efforts [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis limitation has come into focus in recent years. A May 2025 investigative report by Reuters revealed the presence of rogue wireless modules embedded in Chinese-manufactured solar inverters deployed across U.S. infrastructure [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. These undocumented components were invisible to conventional inspection and enabled covert remote access that bypassed all software protections. A similar case, the widely discussed \u0026ldquo;Big Hack\u0026rdquo;, involved allegations of tiny unauthorized devices implanted onto server boards to facilitate cyber-espionage [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Such incidents underscore the growing disconnect between digital cybersecurity protections and the physical verification of hardware.\u003c/p\u003e\u003cp\u003eCurrent approaches to [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] trust rely heavily on documentation, sample testing, and presumed supplier integrity, rather than verifying the actual physical materials present on the board [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. This results in a security blind spot: if a component is visually wrong, but electrically equivalent, or placed in an uninspected area, it is likely to go undetected. Rule-based inspection systems and traceability frameworks were never designed to catch these scenarios.\u003c/p\u003e\u003cp\u003eThis paper addresses that gap by introducing a practical, image-based method to perform component-level [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] using the same AOI data already generated during production. Our approach enables scalable, non-invasive detection of unauthorized component [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan additionalcitationids=\"CR19 CR20\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]changes and structural anomalies, providing a critical layer of trust at the physical level, without requiring [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], golden boards, or complete traceability metadata [\u003cspan additionalcitationids=\"CR23\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis work builds on two foundational capabilities previously developed and patented by our team. The first involves visual AI\u0026ndash;based authentication of electronic components using their physical fingerprint, geometry, lead structure, and surface finish, to identify the manufacturer regardless of top markings [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. The second enables contextual decoding of top marking codes to extract part numbers, lot/date codes, and traceability metadata. These methods have been field-proven through deployment on tens of SMT production lines, accumulating over 5\u0026nbsp;billion component inspections [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The current framework integrates and extends these capabilities toward full-board secure hardware assurance.\u003c/p\u003e"},{"header":"2. Method","content":"\u003cp\u003eThe proposed inspection framework is designed to analyze high-resolution images generated by Automated Optical Inspection (AOI) systems [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], enabling component-level verification and anomaly detection through deep learning. It integrates object detection for identifying and localizing all components on a board, semantic segmentation [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]for analyzing component geometry and solder interfaces, and a downstream anomaly classifier that evaluates deviations from expected patterns.\u003c/p\u003e\u003cp\u003eThis architecture allows the system to detect visual anomalies that indicate tampering, substitution, or undocumented rework, with no reliance on electrical measurements, rule-based tolerances, or golden board comparisons. The workflow, outlined in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, includes dataset acquisition, model training, and real-time inspection on new boards.\u003c/p\u003e\u003cp\u003eThe framework operates along two parallel tracks:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eA bottom-up component analysis, where each part is visually fingerprinted based on body shape, leads or balls, logo, top marking, and package features [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e];\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eA top-down layout inspection, where the entire board is evaluated for missing, added, or misplaced components relative to learned spatial patterns.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis dual-layer strategy enables the system to detect both local and global deviations with high sensitivity and flexibility across board types.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Dataset and Training\u003c/h2\u003e\u003cp\u003eTraining data was acquired from two sources:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eClean production boards, imaged inline during assembly with consistent lighting and camera setup.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eDegraded boards, collected from scrap lots and recycling centers, imaged in both cleaned and uncleaned conditions to capture real-world noise: corrosion, oxidation, broken components, dirt, and surface wear.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThe use of multi-domain data ensures generalization across typical production variability and post-deployment board conditions.\u003c/p\u003e\u003cp\u003eThe dataset was expanded using a bootstrapped training strategy [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]:\u003c/p\u003e\u003cp\u003eInitial models trained on pristine images were used to infer component locations and masks on degraded boards. These predictions were manually reviewed and corrected to generate new labeled examples. The updated data was used to retrain the models. This cycle was repeated until detection and segmentation stabilized across pristine and degraded conditions.\u003c/p\u003e\u003cp\u003eThis approach eliminates the need for golden boards or exhaustive pre-labeling and enables the model to handle image noise, occlusion, and structural damage that typically degrade rule-based inspection systems.\u003c/p\u003e\u003cp\u003eThe dataset and model training strategy build on two prior systems developed by our group: (1) a bottom-view authentication engine used to identify component source based on physical characteristics, and (2) a top-view marking decoder trained to read and interpret manufacturer-specific markings. Both systems have been deployed at scale in production and provided the initial models and annotation tools used in this work.\u003c/p\u003e\u003cp\u003eThis multi-domain dataset supports model generalization across pristine, degraded, and field-recovered board conditions, critical for real-world deployment where production variability and contamination are common.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Object Detection\u003c/h2\u003e\u003cp\u003eComponent-level object detection is performed using the YOLOv12 architecture [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. This model was chosen for its strong performance in dense scenes and its speed, enabling inference under 1000 ms per board. The detector is trained to localize and classify components by package type using CAD-aligned bounding boxes. Data augmentation techniques, flipping, brightness shifts, mild rotation, and Gaussian blu, were applied to improve generalization across lighting and contamination conditions.\u003c/p\u003e\u003cp\u003eIn deployment, the model outputs bounding boxes, class labels, and confidence scores for each component. These detections are used both for component fingerprinting and for guiding segmentation and anomaly classification.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3. Component Segmentation\u003c/h2\u003e\u003cp\u003eA semantic segmentation model complements detection by producing pixel-level masks for each component body and its solder pad regions. These masks enable analysis of footprint alignment, pad coverage, and geometric consistency with reference designs. As with detection, segmentation was trained using both pristine and degraded boards, bootstrapped through iterative retraining. This ensures that the model performs reliably even under noise and partial occlusion.\u003c/p\u003e\u003cp\u003eThe output masks serve two purposes:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eGeometry extraction for anomaly classification\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eLayout validation against expected pad/footprint shapes\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis module enhances the system\u0026rsquo;s ability to detect subtle anomalies such as footprint mismatch, rework, or soldering defects.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eOutput from the semantic segmentation model showing detailed masking of each component\u0026rsquo;s body (shown in blue) and the associated solder pad regions (highlighted in red). These masks enable fine-grained geometric comparisons with the golden reference.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4. Anomaly Classification\u003c/h2\u003e\u003cp\u003eAn anomaly classifier receives features from the detection and segmentation stages. Each component is represented by a feature vector describing:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eBounding box size and aspect ratio\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePad and body mask shape metrics\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eRelative position and spacing\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTexture descriptors and marking alignment\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese features are processed by a shallow CNN that assigns each part to one of several anomaly classes: missing, added, rotated, piggybacked, contaminated, unknown, or reworked. The classifier operates independently of detection confidence, allowing detection and classification decisions to be decoupled.\u003c/p\u003e\u003cp\u003eTo reduce false positives, low-confidence detections and geometric outliers are filtered or flagged. Spatial thresholds are learned from reference layouts and adjusted for each board type.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5. Framework Overview\u003c/h2\u003e\u003cp\u003eThe full system operates as a modular pipeline:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eHigh-resolution AOI images are processed by the object detector.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eDetected regions are passed to the segmentation module.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eCombined features are evaluated by the anomaly classifier.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eOutput anomalies are categorized and scored for downstream action (alerting, review, traceability linking, etc.).\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eThe system operates at production speed and requires no modification to the AOI hardware or imaging process. Once trained, it performs inspection autonomously, providing board-level results in sub-second timeframes.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Experimental Results and Case Studies","content":"\u003cp\u003eTo evaluate the performance of the proposed AOI image\u0026ndash;based inspection framework, we conducted experiments across diverse datasets representing both clean production conditions and degraded, real-world boards. The evaluation focused on component detection accuracy, segmentation quality, anomaly classification, and overall robustness under varying visual conditions.\u003c/p\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Detection and Segmentation Performance\u003c/h2\u003e\u003cp\u003eThe object detection model, based on YOLOv12, was trained using three dataset types:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003ePristine AOI images, aligned with CAD data;\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eCleaned scrap boards, scanned under controlled lighting;\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eUncleaned scrap boards, representing worst-case conditions including oxidation, broken parts, and contamination.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eModel training followed the bootstrapped strategy described in Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Training convergence over 360 epochs is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, which illustrates loss stability and generalization across the full dataset.\u003c/p\u003e\u003cp\u003eModel performance on unseen boards is summarized in Table I. The model achieved an average F1 score of 0.90 across board types, with precision and recall remaining high even on degraded inputs.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eTable I \u0026ndash;\u003c/b\u003e Detection Performance Across Board Conditions.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"4\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBoard Condition\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eF1 Score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePristine AOI Image\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClean Scrap Board\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDirty Scrap Board\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAverage\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.90\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.88\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.94\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThese results confirm the detector\u0026rsquo;s resilience. Although performance dips under severe degradation, precision and recall remain strong, enabling reliable component localization for downstream analysis.\u003c/p\u003e\u003cp\u003eThe segmentation model was evaluated on 2,371 component instances across 241 AOI images. Results are shown in Table II. High mAP50 scores demonstrate accurate mask alignment, particularly on solder pads.\u003c/p\u003e\u003cp\u003e\u003cb\u003eTable II \u0026ndash;\u003c/b\u003e Segmentation Accuracy by Class.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e\u003ccolgroup cols=\"4\"\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\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClass\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\u003emAP50\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBody\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.911\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.919\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.961\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePad\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.975\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.934\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.978\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAll\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.943\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.927\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.969\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. \u003cem\u003ePrecision-Confidence Curve for Segmentation.\u003c/em\u003e The model maintains stable behavior across confidence thresholds, with average segmentation precision around 0.92.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Real-World Case Studies\u003c/h2\u003e\u003cp\u003eThe framework was applied to production boards to evaluate its effectiveness in detecting undocumented deviations. These real-world examples demonstrate its ability to catch anomalies that would bypass conventional traceability and AOI systems.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eIn one case, a microcontroller was flagged for inconsistent markings. Visual comparison revealed it was sourced from a different manufacturer, despite electrical equivalence. The system flagged the anomaly based on shape and logo inconsistencies.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eAnother board contained a passive component logged as Panasonic, but identified visually as a Vishay part. Package geometry and finish mismatches led to its classification as a substitution.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eA separate example involved a part logged as Bourns but unrecognized by the model. The system classified it as unknown, and manual inspection confirmed it was a counterfeit or unqualified device.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e shows layout-level anomalies detected through top-down board inspection. Components added, missing, or misplaced relative to the learned layout are clearly flagged.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe system also detected piggybacking\u0026mdash;a technique where unauthorized components are soldered atop legitimate ones to alter function. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows detection of such tampering on a board, where geometric irregularities triggered anomaly classification.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn another case, a wire had been manually soldered between two nodes post-production, likely to override circuit behavior or inject a side-channel function. Figure\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e shows this anomaly, along with an obscuring sticker, both caught by layout-based anomaly detection.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFinally, Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e shows a component substitution where a crystal oscillator was visually replaced with an alternate vendor\u0026rsquo;s part. The substitution was detected via shape, pad layout, and silkscreen misalignment.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Scalability and Throughput\u003c/h2\u003e\u003cp\u003ePerformance was benchmarked across full board scans at standard AOI resolutions. The system achieves:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u0026gt;\u0026thinsp;99.3% anomaly detection accuracy\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.5% false positive rate\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026lt;\u0026thinsp;1% false negative rate\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u0026lt;\u0026thinsp;1000 ms total processing time per board\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eComponent-level inference runs at under 10 ms per instance. This enables full-coverage inspection without disrupting throughput or requiring manual review. The framework integrates directly into post-AOI workflows and can also be applied retrospectively for forensics or supply chain audits.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThis work introduces a novel approach to hardware assurance: a visual AI system that performs full-coverage component-level verification using only AOI image data. Unlike conventional systems, which rely on rule-based logic, golden boards, or traceability records, our method directly analyzes the visual identity and spatial context of every component on the board, enabling the detection of anomalies that were previously undetectable at scale.\u003c/p\u003e\u003cp\u003eThe framework described here is a direct evolution of two earlier, production-grade systems that separately addressed component-level authentication and marking interpretation. By unifying these capabilities into a board-level inspection architecture and training on a much broader dataset, we demonstrate, for the first time, scalable, component-by-component hardware assurance directly from AOI imagery.\u003c/p\u003e\u003cp\u003eThe core novelty lies in the dual-layer inspection strategy:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eA bottom-up component fingerprinting process, where each part is analyzed individually for shape, lead configuration, markings, and logos; and\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eA top-down board-level consistency check, which learns spatial relationships across multiple known-good boards and flags layout anomalies such as missing, swapped, or added components.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eTogether, these layers allow the system to detect both local and global modifications, including unauthorized substitutions, rework, and tampering, without requiring any electrical testing or metadata. To our knowledge, no prior work has demonstrated this level of scalable, image-only hardware verification on production boards.\u003c/p\u003e\u003cp\u003eThe system\u0026rsquo;s generalization ability is another key contribution. Through a bootstrapped training pipeline, it learns from both pristine and degraded board conditions, extending its performance beyond lab settings to real-world factory floors, RMA returns, or field-recovered units. This enables the method to maintain high accuracy even in the presence of oxidation, contamination, or partial damage, scenarios that defeat traditional rule-based AOI and even most learning-based defect classifiers.\u003c/p\u003e\u003cp\u003eAdditionally, the use of standard AOI imagery makes the method immediately deployable. It introduces no production bottlenecks and requires no changes to inspection equipment. This enables manufacturers to repurpose their existing image streams, not only for quality assurance but also for supply chain verification, provenance validation, and hardware security enforcement.\u003c/p\u003e\u003cp\u003eWhile performance is strong, several practical limitations remain. The system depends on sufficient imaging quality, which can vary across AOI platforms. It also cannot determine intent, e.g., whether an anomaly stems from malicious tampering, unapproved rework, or an upstream substitution error. Expanding the component reference database and improving automated anomaly attribution will be key for scaling across product lines and sectors.\u003c/p\u003e\u003cp\u003eNonetheless, this framework represents a significant evolution in how manufacturing data can be used. By combining object detection, semantic segmentation, and anomaly classification in a modular pipeline, and grounding inspection in visual pattern learning rather than rule definition, the system provides a new pathway toward trusted electronics at the physical layer.\u003c/p\u003e\u003cp\u003eThis methodology repositions AOI imagery from a passive QA artifact to a forensic-grade security tool, enabling industries to verify not only whether components were placed, but whether the right components, from the right sources, were used.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis work presents a novel, scalable framework for secure hardware assurance based entirely on AOI image data and deep learning. By analyzing each component\u0026rsquo;s visual identity, including shape, lead structure, and markings, and comparing board-wide layouts across known-good samples, the system detects subtle anomalies such as substitutions, unauthorized additions, or tampering, without the need for electrical testing or traceability metadata.\u003c/p\u003e\u003cp\u003eUnlike existing inspection or authentication methods, which are limited by rules, golden boards, or metadata availability, our approach performs autonomous, full-coverage verification using data already captured in standard manufacturing workflows. This transforms AOI from a passive quality assurance step into an active layer of physical trust enforcement.\u003c/p\u003e\u003cp\u003eThe method has been validated across a broad range of conditions, from pristine boards to visually degraded and field-returned samples, demonstrating high accuracy, generalization, and fast inference suitable for in-line deployment. Its ability to detect deviations that evade conventional AOI or software-based security makes it well-suited also for critical sectors such as aerospace, defense, and medical electronics.\u003c/p\u003e\u003cp\u003eBy bridging the gap between image-based inspection and hardware cybersecurity, this system offers a new model for defending the physical supply chain, where hardware authenticity is no longer assumed, but verified.\u003c/p\u003e\u003cp\u003eThis framework integrates two previously patented systems: one for identifying component origin via visual fingerprinting, and another for decoding manufacturer-specific top markings. Both have been deployed across major production lines and trained on over 5\u0026nbsp;billion AOI images. By combining these proven capabilities with board-level spatial reasoning and anomaly detection, the presented solution delivers a new standard for hardware trust at the physical layer, scalable, autonomous, and grounded in real-world manufacturing data.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthical Approval\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003enot applicable\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003enot applicable\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis study does not rely on publicly available datasets. All data generated or analyzed during this study are either proprietary or derived from internal operational systems and are not publicly available due to confidentiality agreements. Reasonable requests for access to anonymized data supporting the findings of this study may be considered by the corresponding author, subject to approval and compliance with relevant data protection regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEyal Weiss conceived the research, developed the methodology, conducted the experiments, prepared the figures, and wrote the manuscript. The author reviewed and approved the final version.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eD. Mehta, H. Lu, O. P. Paradis, M. A. MS, M. T. Rahman, Y. Iskander, P. Chawla, D. L. Woodard, M. Tehranipoor, and N. Asadizanjani, ACM Journal on Emerging Technologies in Computing Systems (JETC) \u003cstrong\u003e16\u003c/strong\u003e, 1 (2020).\u003c/li\u003e\n\u003cli\u003eM. M. Tehranipoor, U. Guin, and D. Forte, in \u003cem\u003eCounterfeit Integrated Circuits\u003c/em\u003e (Springer, 2015), pp. 15\u0026ndash;36.\u003c/li\u003e\n\u003cli\u003eS. Akter, K. Khalil, and M. Bayoumi, IEEe Access \u003cstrong\u003e11\u003c/strong\u003e, 77543 (2023).\u003c/li\u003e\n\u003cli\u003eH. Representatives, (2017).\u003c/li\u003e\n\u003cli\u003eM. Tehranipoor, N. Pundir, N. Vashistha, and F. Farahmandi, \u003cem\u003eHardware Security Primitives\u003c/em\u003e (Springer, 2023).\u003c/li\u003e\n\u003cli\u003eR. Ajax, (n.d.).\u003c/li\u003e\n\u003cli\u003eA. Hemmati, P. Raoufi, and A. M. Rahmani, Neural Comput Appl \u003cstrong\u003e36\u003c/strong\u003e, 11461 (2024).\u003c/li\u003e\n\u003cli\u003eM. M. M. Islam, J. I. Emon, K. Y. Ng, A. Asadpour, M. M. R. Al Aziz, M. L. Baptista, and J.-M. Kim, in \u003cem\u003eArtificial Intelligence for Smart Manufacturing and Industry X. 0\u003c/em\u003e (Springer, 2025), pp. 9\u0026ndash;36.\u003c/li\u003e\n\u003cli\u003eV. Reshadat and R. A. J. W. Kapteijns, in \u003cem\u003e2021 International Conference on Data and Software Engineering (ICoDSE)\u003c/em\u003e (IEEE, 2021), pp. 1\u0026ndash;5.\u003c/li\u003e\n\u003cli\u003eL. Meyer, B. Josefsson, K. Vrotsou, C. Westin, and J. Lundberg, in \u003cem\u003e2021 Ieee/Aiaa 40th Digital Avionics Systems Conference (Dasc)\u003c/em\u003e (IEEE, 2021), pp. 1\u0026ndash;8.\u003c/li\u003e\n\u003cli\u003eX. Tong, Z. Yu, X. Tian, H. Ge, and X. Wang, Front Comput Sci \u003cstrong\u003e16\u003c/strong\u003e, 1 (2022).\u003c/li\u003e\n\u003cli\u003eY.-S. Ni, W.-L. Chen, Y. Liu, M.-H. Wu, and J.-I. Guo, Sensors \u003cstrong\u003e24\u003c/strong\u003e, 5737 (2024).\u003c/li\u003e\n\u003cli\u003eI.-C. Chen, R.-C. Hwang, and H.-C. Huang, Processes \u003cstrong\u003e11\u003c/strong\u003e, 775 (2023).\u003c/li\u003e\n\u003cli\u003eSarah Mcfarlane, Reuters (2025).\u003c/li\u003e\n\u003cli\u003eA. Spieske and H. Birkel, Comput Ind Eng \u003cstrong\u003e158\u003c/strong\u003e, 107452 (2021).\u003c/li\u003e\n\u003cli\u003eS. P. Gayialis, E. P. Kechagias, G. A. Papadopoulos, and D. Masouras, Sustainability (Switzerland) \u003cstrong\u003e14\u003c/strong\u003e, (2022).\u003c/li\u003e\n\u003cli\u003eM. T. Rahman, Q. Shi, S. Tajik, H. Shen, D. L. Woodard, M. Tehranipoor, and N. Asadizanjani, in \u003cem\u003e2018 IEEE 3rd International Verification and Security Workshop (IVSW)\u003c/em\u003e (2018), pp. 93\u0026ndash;102.\u003c/li\u003e\n\u003cli\u003eM. M. Tehranipoor, U. Guin, and D. Forte, in \u003cem\u003eCounterfeit Integrated Circuits\u003c/em\u003e (Springer, 2015), pp. 95\u0026ndash;107.\u003c/li\u003e\n\u003cli\u003eM. M. Tehranipoor, U. Guin, and D. Forte, in \u003cem\u003eCounterfeit Integrated Circuits\u003c/em\u003e (Springer, 2015), pp. 75\u0026ndash;93.\u003c/li\u003e\n\u003cli\u003eL. W. Kessler and T. Sharpe, Printed Circuit Design \u0026amp; Fab \u003cstrong\u003e27\u003c/strong\u003e, 64 (2010).\u003c/li\u003e\n\u003cli\u003ea S. Report, T (2011).\u003c/li\u003e\n\u003cli\u003eG. M. Razak, L. C. Hendry, and M. Stevenson, Production Planning and Control \u003cstrong\u003e34\u003c/strong\u003e, 1114 (2023).\u003c/li\u003e\n\u003cli\u003eP. Rosales, (2021).\u003c/li\u003e\n\u003cli\u003eR. Schuitemaker and X. Xu, in \u003cem\u003eProcedia CIRP\u003c/em\u003e (Elsevier B.V., 2020), pp. 700\u0026ndash;705.\u003c/li\u003e\n\u003cli\u003eE. Weiss, IEEE Trans Compon Packaging Manuf Technol 1 (2024).\u003c/li\u003e\n\u003cli\u003eE. Weiss, SMTA (2021).\u003c/li\u003e\n\u003cli\u003eE. Weiss and Z. Efrat, P-603538-PC (2021).\u003c/li\u003e\n\u003cli\u003eY. Li, Y. Duan, L. Duan, W. Xiang, and Q. Wu, in \u003cem\u003eMedical Image Understanding and Analysis\u003c/em\u003e, edited by M. H. Yap, C. Kendrick, A. Behera, T. Cootes, and R. Zwiggelaar (Springer Nature Switzerland, Cham, 2024), pp. 143\u0026ndash;159.\u003c/li\u003e\n\u003cli\u003eJ.-B. Grill, F. Strub, F. Altch\u0026eacute;, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Daniel Guo, M. Gheshlaghi Azar, B. Piot, K. Kavukcuoglu, R. Munos, and M. Valko, \u003cem\u003eBootstrap Your Own Latent A New Approach to Self-Supervised Learning\u003c/em\u003e (n.d.).\u003c/li\u003e\n\u003cli\u003eY. Li, S. Li, H. Du, L. Chen, D. Zhang, and Y. Li, IEEE Access \u003cstrong\u003e8\u003c/strong\u003e, 227288 (2020).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-hardware-and-systems-security","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hass","sideBox":"Learn more about [Journal of Hardware and Systems Security](https://link.springer.com/journal/41635)","snPcode":"41635","submissionUrl":"https://submission.nature.com/new-submission/41635/3","title":"Journal of Hardware and Systems Security","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"visual inspection, anomaly detection, secure hardware assurance, deep learning, automated optical inspection (AOI), component authentication, PCB assem-bly, counterfeit detection, manufacturing integrity, hardware cybersecurity","lastPublishedDoi":"10.21203/rs.3.rs-7197469/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7197469/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEnsuring the authenticity and integrity of electronic assemblies is increasingly critical as hardware-based attacks and unauthorized component modifications become more sophisticated. Conventional inspection systems, whether rule-based AOI or traceability logs, offer limited protection against subtle or intentional tampering.\u003c/p\u003e\u003cp\u003eThis paper introduces a deep learning\u0026ndash;based framework for secure hardware assurance that operates directly on AOI image data, enabling autonomous, full-coverage verification of every component on the board. The method is built on two previously patented systems: one for component authentication via visual fingerprinting\u0026mdash;independent of top marking\u0026mdash;and another for contextual decoding of top marking codes. These systems have been deployed across tens of SMT lines, generating over 5\u0026nbsp;billion production-grade inspections.\u003c/p\u003e\u003cp\u003eBy integrating and extending these capabilities, the system performs bottom-up part analysis and top-down layout validation, identifying substitutions, rework, and tampering, without requiring electrical probing, golden boards, or metadata.\u003c/p\u003e\u003cp\u003eResults show\u0026thinsp;\u0026gt;\u0026thinsp;99% detection accuracy and sub-second inspection times, enabling secure, image-only verification suitable for in-line or forensic use. This work bridges the gap between software protections and physical hardware trust, transforming AOI images into a practical security and compliance enforcement tool.\u003c/p\u003e","manuscriptTitle":"Secure Hardware Assurance Using Visual AI on AOI Imaging of Electronic Assemblies","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-06 08:50:31","doi":"10.21203/rs.3.rs-7197469/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-10-05T16:07:39+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-28T02:35:53+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"171380182092872428097615145141448239137","date":"2025-08-09T10:48:27+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-07T07:09:22+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"330017666678469478453874038004621794050","date":"2025-08-05T02:54:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"80258550166170316101286447436859554211","date":"2025-08-04T11:58:00+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-04T10:53:56+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-04T03:14:47+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-01T00:29:11+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Hardware and Systems Security","date":"2025-07-23T14:24:53+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-hardware-and-systems-security","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hass","sideBox":"Learn more about [Journal of Hardware and Systems Security](https://link.springer.com/journal/41635)","snPcode":"41635","submissionUrl":"https://submission.nature.com/new-submission/41635/3","title":"Journal of Hardware and Systems Security","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"1bed5aa2-a4d5-496e-8815-201f4825620b","owner":[],"postedDate":"August 6th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-16T16:03:48+00:00","versionOfRecord":{"articleIdentity":"rs-7197469","link":"https://doi.org/10.1007/s41635-026-00173-5","journal":{"identity":"journal-of-hardware-and-systems-security","isVorOnly":false,"title":"Journal of Hardware and Systems Security"},"publishedOn":"2026-03-09 15:59:09","publishedOnDateReadable":"March 9th, 2026"},"versionCreatedAt":"2025-08-06 08:50:31","video":"","vorDoi":"10.1007/s41635-026-00173-5","vorDoiUrl":"https://doi.org/10.1007/s41635-026-00173-5","workflowStages":[]},"version":"v1","identity":"rs-7197469","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7197469","identity":"rs-7197469","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}