Paper deep dive
A Benchmark & Dataset for Detecting AI-Manipulated Visual Evidence in the Court System
Kelly McConvey, Sajad Ebrahimi, Nima Jamali, Jalehsadat Mahdavimoghaddam, Matina Mahdizadeh Sani, Maksym Taranukhin, Wentao Zhang, Jacquelyn Burkell, Yuntian Deng, Karen Eltis, Maura R. Grossman, Vered Shwartz, Ebrahim Bagheri
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Photographic evidence is becoming increasingly vulnerable to forms of alteration and fabrication that existing legal and technical workflows are not well equipped to evaluate. Surveillance frames, dashcam stills, and phone photographs may be used to establish presence, sequence, causation, damage, or identity, yet contemporary generative systems allow non-experts to alter or fabricate such images through ordinary prompt-based interfaces. Existing image-forensics benchmarks provide important resources for face manipulation, classical tampering, and general synthetic-image detection, but they are not organized around the forms of visual evidence submitted in courts, the localized edits that can change what an exhibit appears to prove, or the consumer-tool threat model now facing the justice system. We introduce the CIFAR Synthetic Evidence Corpus for Detecting AI-Manipulated Images, a benchmark for evidentiary image authentication in court and justice-system contexts. The corpus contains 1,505 photographic items, including 720 authentic controls and 785 manipulated or fabricated images, spanning surveillance, dashcam, and consumer-photo imagery. Manipulations are organized into scene-condition edits, localized element edits, and full fabrications produced with contemporary generative systems. Each item is released with structured metadata covering source provenance, manipulation tier, subtype, generator, prompt template, and scene attributes, enabling controlled evaluation beyond aggregate binary detection. We also establish baselines with publicly available image-manipulation detectors, showing that current systems exhibit error profiles that remain problematic for evidentiary use. The dataset, prompts, metadata, code, and baseline evaluation scripts are released to support research on visual evidence authentication, information integrity, and trustworthy AI for the justice system.
Tags
Links
- Source: https://arxiv.org/abs/2609.37783v1
- Canonical: https://arxiv.org/abs/2609.37783v1
Trouble viewing inline? Open PDF directly →
Full Text
29,344 characters extracted from source content.
Expand or collapse full text
A Benchmark & Dataset for Detecting AI-Manipulated Visual Evidence in the Court System Conference: Proceedings of the 35th ACM International Conference on Information and Knowledge Management; November 7–11, 2026; Rome, Italy.Proceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM ’26), November 7–11, 2026, Rome, ItalyISBN: 979-8-4007-2539-5/2026/11DOI: 10.1145/3799682.3840194CCS: Computing methodologies Computer visionCCS: Applied computing LawCCS: Information systems Multimedia and multimodal retrieval Kelly McConvey Note: Corresponding author. Email: kelly.mcconvey@mail.utoronto.ca Affiliation: University of Toronto, Toronto, Canada , Sajad Ebrahimi Affiliation: University of Toronto, Toronto, Canada , Nima Jamali Affiliation: University of Waterloo, Waterloo, Canada , Jalehsadat Mahdavimoghaddam Affiliation: University of Toronto, Toronto, Canada , Matina Mahdizadeh Sani Affiliation: University of Waterloo, Waterloo, Canada , Maksym Taranukhin Affiliation: University of British Columbia, Vancouver, Canada , Wentao Zhang Affiliation: University of Waterloo, Waterloo, Canada , Jacquelyn Burkell Affiliation: University of Western Ontario, London, Canada , Yuntian Deng Affiliation: University of Waterloo, Waterloo, Canada , Karen Eltis Affiliation: University of Ottawa, Ottawa, Canada , Maura R. Grossman Affiliation: University of Waterloo, Waterloo, Canada , Vered Shwartz Affiliation: University of British Columbia, Vancouver, Canada and Ebrahim Bagheri Affiliation: University of Toronto, Toronto, Canada © c Abstract. Photographic evidence is becoming increasingly vulnerable to forms of alteration and fabrication that existing legal and technical workflows are not well equipped to evaluate. Surveillance frames, dashcam stills, and phone photographs may be used to establish presence, sequence, causation, damage, or identity, yet contemporary generative systems allow non-experts to alter or fabricate such images through ordinary prompt-based interfaces. Existing image-forensics benchmarks provide important resources for face manipulation, classical tampering, and general synthetic-image detection, but they are not organized around the forms of visual evidence submitted in courts, the localized edits that can change what an exhibit appears to prove, or the consumer-tool threat model now facing the justice system. We introduce the CIFAR Synthetic Evidence Corpus for Detecting AI-Manipulated Images, a benchmark for evidentiary image authentication in court and justice-system contexts. The corpus contains 1,505 photographic items, including 720 authentic controls and 785 manipulated or fabricated images, spanning surveillance, dashcam, and consumer-photo imagery. Manipulations are organized into scene-condition edits, localized element edits, and full fabrications produced with contemporary generative systems. Each item is released with structured metadata covering source provenance, manipulation tier, subtype, generator, prompt template, and scene attributes, enabling controlled evaluation beyond aggregate binary detection. We also establish state-of-the-art baselines with publicly available image-manipulation detectors, showing that current systems exhibit error profiles that remain problematic for evidentiary use. The dataset, prompts, metadata manifest, code, and baseline evaluation scripts are released to support research on visual evidence authentication, information integrity, and trustworthy AI for the justice system. Keywords: image forensics, evidentiary authentication, benchmark dataset, AI-generated content, visual evidence, deepfake detection †c-license: by-nc-nd 1. Introduction Courts increasingly receive photographs whose authenticity can no longer be taken for granted, including surveillance frames offered to show that a person was present at a location, dashcam stills used to reconstruct traffic incidents, and phone photographs submitted to document damage, injury, possession, or the condition of a scene. In each case, the image is not merely visual content, but evidence whose value depends on whether the depicted scene can be trusted, whether relevant details have been altered, and whether the image can be situated within a reliable chain of provenance. This paper addresses that problem as one of evidentiary image authentication, where the task is to evaluate whether photographic evidence is authentic, manipulated, or synthetic under conditions that resemble legal and institutional use. Photographs have never been self-authenticating records of reality, since courts have long required them to be connected to testimony, context, and procedure before they can function as evidence (Grossman and Grimm, 2025). Historical accounts of photographic admissibility show that courts have repeatedly had to reconcile the apparent objectivity of photographs with the possibility of staging, selection, distortion, and later digital alteration (Mnookin, 1998). The current challenge is therefore not that visual evidence has suddenly become uncertain, but that the practical conditions under which courts have managed that uncertainty have changed. Evidence that once required specialized skill to fabricate can now be altered or produced through widely available generative systems. Fabricated and misleading evidence is not new, yet generative AI changes its cost, accessibility, and plausibility in ways that are directly relevant to evidentiary workflows. Earlier forms of photographic manipulation often required specialized tools, technical expertise, or substantial effort, whereas contemporary image-generation systems allow non-experts to produce plausible visual alterations through ordinary prompt-based interfaces. A litigant, witness, or third party can change scene conditions, insert or remove salient objects, or generate a plausible image from scratch without the expertise that previously constrained visual forgery. The same technological shift also creates a second evidentiary risk, since genuine evidence may be dismissed as fabricated precisely because fabrication itself has become credible. This is the dynamic Chesney and Citron (Chesney and Citron, 2018) describe as the liar’s dividend, and it places pressure on courts and related institutions not only to detect false evidence, but also to preserve confidence in authentic evidence. Recent legal developments show that this concern has moved beyond speculation, as courts and judicial organizations are now confronting acknowledged and suspected AI-generated material, including fabricated legal authorities, synthetic media, and allegedly manipulated exhibits. In Mata v. Avianca (Ryan and Garrett, 2023), generative AI produced plausible but non-existent legal authorities that were submitted in litigation, illustrating how synthetic material can enter formal legal processes when verification fails. More directly for visual evidence, judicial and legal scholarship has begun to examine how courts should handle acknowledged and unacknowledged AI-generated evidence, while court-focused organizations have warned that synthetic evidence may affect public trust in judicial proceedings (Heaton et al., 2026). These developments make evidentiary image authentication a timely problem for legal institutions and for computational researchers working on information integrity, provenance, verification, and trustworthy decision support. The evidentiary risk is not limited to fully synthetic images, because in many legal settings the more consequential manipulation is a localized or contextual edit to an otherwise ordinary photograph. A person may be removed from a surveillance frame, a road sign or traffic signal may be changed in a dashcam still, or visible damage may be added to or removed from a phone photograph. Such edits can alter what an image appears to prove while preserving most of its visual content, which makes them difficult for lay observers (Bray et al., 2023; Roca et al., 2025) and also difficult for detectors trained mainly on whole-image synthesis or face-centered deepfakes. For evidentiary authentication, the relevant question is therefore not simply whether an image was generated by AI, but whether verification systems can identify the forms of manipulation that change the informational and evidentiary meaning of a visual item. Existing datasets do not adequately support this domain. Face-deepfake corpora have enabled important progress in synthetic-media detection, but they primarily address identity, expression, and facial manipulation rather than scene-level photographic evidence. General image-tampering datasets capture classical operations such as splicing, copy-move, and removal, but many predate prompt-driven consumer generation and do not reflect the artifacts, workflows, or editing patterns now available to non-experts. Other public resources provide authentic surveillance, driving, or consumer imagery, yet they do not provide controlled AI-generated manipulations, prompt metadata, generator metadata, or benchmark-ready labels. What is missing is a public resource for evaluating authentication systems on evidentiary image families, realistic manipulation types, consumer-interface generation, structured provenance, and detector behavior under conditions relevant to institutional use. We introduce the CIFAR Synthetic Evidence Corpus for Detecting AI-Manipulated Images, a dataset and benchmark for evidentiary image authentication. The corpus contains 1,505 photographic items, including 720 authentic controls and 785 manipulated or fabricated images. It spans three families of visual evidence, namely surveillance imagery, dashcam stills, and consumer photographs. Manipulations are organized into three tiers covering scene-condition changes, localized element edits, and complete fabrication. The manipulated images are produced using contemporary generative systems from multiple organizations, and each item is paired with metadata describing its source provenance, evidentiary family, manipulation tier, generator, prompt template, and scene attributes. This paper makes four contributions. First, it releases a benchmark corpus of authentic and AI-manipulated photographic evidence across surveillance, dashcam, and consumer-photo imagery. Second, it introduces a manipulation taxonomy grounded in evidentiary practice, with particular attention to localized edits that can alter the meaning of a visual exhibit while leaving most of the scene intact. Third, it provides a generation pipeline based on a realistic non-expert consumer-tool threat model. Fourth, it reports a zero-shot benchmark of off-the-shelf detectors, showing that current systems exhibit error profiles that remain problematic for evidentiary use. The corpus is available at https://github.com/UofT-CIFAR/Synthetic-Evidence-Image-Corpus, DOI: 10.5281/zenodo.22015105. 2. Distinction from Existing Resources Our corpus complements rather than replaces existing image-forensics benchmarks by targeting the authentication of photographic evidence under contemporary AI-enabled manipulation. Face-manipulation corpora such as FaceForensics++, Celeb-DF, DFDC, ForgeryNet, and OpenForensics (Rössler et al., 2019; Li et al., 2020; Dolhansky et al., 2020; He et al., 2021; Le et al., 2021) primarily treat the face as the object of authentication, whereas our corpus treats the photograph as a visual claim about an event, location, object, or condition. General tampering benchmarks such as CASIA and the NIST Media Forensics Challenge datasets (Dong et al., 2013; Guan et al., 2019) provide important coverage of splicing, copy-move, removal, and related operations, but largely reflect editing paradigms that predate consumer generative systems. Our corpus instead emphasizes prompt-driven generation and editing, including localized changes, scene-level modifications, and complete fabrication, capturing a contemporary non-expert threat model (Corvi et al., 2022; Chandra et al., 2025). Similarly, authentic-image resources such as VIRAT, UCF-Crime, BDD100K, and CASIA (Oh et al., 2011; Sultani et al., 2019; Yu et al., 2020; Dong et al., 2013) provide valuable source imagery but do not pair it with controlled generative edits, generation metadata, and manipulation labels designed for authentication analysis. Our corpus extends this ecosystem by organizing manipulated photographic evidence according to evidentiary family, manipulation tier and subtype, source dataset, generator, prompt, and scene attributes. This structure supports analysis beyond aggregate detection accuracy, allowing researchers to evaluate which evidentiary changes are detectable, which are missed, and which authentic images generate false positives. This focus addresses a gap identified by legal scholarship, which has increasingly emphasized the need for reliable authentication of AI-generated evidence and stronger mechanisms for addressing synthetic-media risks in litigation (Chesney and Citron, 2018; Delfino, 2023; Dalal et al., 2025; Grimm et al., 2021; Grossman and Grimm, 2025). While that literature establishes the institutional problem, it does not provide the empirical infrastructure needed to measure authentication performance on evidentiary photographs. Our corpus addresses this missing empirical layer without claiming that detection alone can resolve the legal problem; instead, it provides a benchmark for studying the reliability, failure modes, and limits of authentication systems in a setting aligned with evidentiary practice. 3. Resource Design and Construction (a) Authentic source (b) T1: weather edit (c) Authentic source (d) T2: person added Figure 1. Example manipulations. Top: Tier-1 scene-condition edit adds snow to a dashcam frame. Bottom: Tier-2 localized edit inserts a person into a surveillance frame.Four image exemplars from the dataset. The top two images depict before and after of Tier-1 scene-condition edit adds snow to a dashcam frame. The bottom two images depict before and after of Tier-2 localized edit inserts a person into a surveillance frame. Our proposed corpus is designed for evidentiary image authentication rather than general synthetic-image detection. This setting requires a benchmark that represents the kinds of photographs submitted in legal and institutional contexts, includes manipulations that can change the evidentiary meaning of an image, reflects the consumer-tool threat model through which non-experts can now produce plausible edits, and provides metadata that supports controlled analysis beyond aggregate binary classification. We operationalize these requirements along four axes. First, the corpus covers three evidentiary families: surveillance, dashcam, and consumer photographs. Second, it distinguishes three manipulation tiers: scene-condition changes (T1), localized element edits (T2), and complete fabrication (T3); Table 1 reports the full subtype breakdown by family. Figure 1 shows representative T1 and T2 manipulations. Third, we use consumer-accessible generative systems from multiple organizations to reduce dependence on artifacts from a single provider. Fourth, each item includes metadata describing its source, evidentiary family, manipulation tier and subtype, generator, prompt, and scene attributes. For constructing the corpus, we first defined a fixed set of prompt templates, each tied to a family, tier, and subtype. The templates were written to model a non-expert using a consumer image generation interface. Each prompt requests a single scoped change, asks the model to preserve all other scene content, and aims to produce a plausible evidentiary image rather than an arbitrary edit. The final prompt set contains 27 templates spanning the three families and three tiers. We constructed the source pool from public datasets matched to the three evidentiary families: VIRAT and UCF-Crime for surveillance, BDD100K for dashcam imagery, and CASIA v2.0 for consumer photographs (Oh et al., 2011; Sultani et al., 2019; Yu et al., 2020; Dong et al., 2013). Tier 1 and Tier 2 manipulations use authentic source frames, while Tier 3 images are generated directly from prompts. Candidate frames were filtered using automatically extracted scene attributes to satisfy predefined quotas, and no source frame is reused. Table 1. Manipulated image corpus by evidentiary family, source dataset, and manipulation tier. Family Source dataset(s) T1 (scene conditions) T2 (element edit) T3 Total Surveillance VIRAT, UCF-Crime Time of day 30 Swap object 28 45 264 Weather 28 Remove person 26 Lighting 29 Remove signage 25 Crowd density 30 Add person 23 Dashcam BDD100K Time of day 29 Remove sign 30 45 255 Weather 27 Traffic-light color 27 Lighting 29 Remove vehicle 27 Crowd density 27 Add hazard 14 Consumer photos CASIA Time of day 30 Add object 29 45 266 Weather 30 Remove damage 28 Lighting 30 Swap object 25 Crowd density 30 Add damage 19 Total 349 301 135 785 Manipulated items were generated by applying the appropriate prompt template to the assigned source frame for Tier 1 and Tier 2, or from the prompt alone for Tier 3. Items were screened during generation and regenerated where the requested edit failed to appear or the scene was visibly degraded; this was not conducted as a formal annotation exercise, so we do not report inter-annotator agreement. The main corpus uses GPT Image 2, Gemini 3.1 with Nano Banana 2, and Firefly 5. GPT Image 2 and Gemini 3.1 were accessed through APIs, while Firefly 5 was accessed through its consumer interface. The corpus also includes a smaller Hero set produced through iterative human-in-the-loop refinement in Photoshop and ChatGPT. The main variants model ordinary single-pass manipulation, while the Hero set models a higher-effort adversary willing to manually refine an image. 4. Dataset Content and Benchmarking Table 2. Performance of off-the-shelf detectors. Detector Acc. Prec. Rec. FPR ROC-AUC PR-AUC d3 0.616 0.696 0.469 0.224 0.674 0.684 univfd 0.573 0.686 0.336 0.168 0.643 0.633 drct 0.535 0.647 0.238 0.141 0.621 0.642 cnnspot 0.478 0.500 0.003 0.003 0.639 0.587 To establish reference points for future work, we evaluate publicly available image-manipulation detectors on the corpus in a zero-shot setting. Each detector scores every image once using its released weights and default decision threshold, with no fine-tuning, and we report both threshold-free metrics (ROC-AUC, PR-AUC), which rank detectors independently of any operating point, and threshold-dependent metrics (accuracy, precision, recall, FPR) at each detector’s shipped default. The evaluation uses the DetectZoo framework (Ebrahimi et al., 2026) and treats the task as binary authentication over all 1,505 items. The detectors span released synthetic-image detectors: a diffusion-reconstruction method (d3), a CLIP-feature classifier (univfd), a diffusion-reconstruction contrastive detector (drct), and a CNN-based GAN-artifact detector (cnnspot). Across all systems, performance is low: the best detector, d3, reaches only 0.674 ROC-AUC, and the rest sit near chance (Table 2). This is the central benchmarking result—off-the-shelf detectors, trained predominantly on fully synthesized GAN or diffusion imagery, transfer poorly to the localized edits and consumer-tool manipulations that characterize evidentiary content. Default operating points are also poorly aligned with evidentiary use: at its shipped threshold d3 catches fewer than half of all manipulations (47% recall) yet wrongly flags 22% of authentic images as manipulated, and falsely impugning genuine evidence is the more damaging error in court. A benchmark for evidentiary authentication must therefore report this error structure, not only aggregate accuracy. A breakdown of the strongest detector, d3, illustrates the diagnostic value of the corpus design. Performance is uneven in ways that track evidentiary structure rather than overall difficulty. By family, dashcam imagery is markedly harder than surveillance or consumer photographs (AUC 0.55 versus 0.69 and 0.78). By generator, detectability varies sharply: items from one consumer model are far less detectable than another’s (per-generator recall ranging from 0.46 to 0.71), confirming that the corpus exposes cross-generator generalization gaps. The high-effort Hero set is the most revealing case: manually refined Photoshop edits receive extremely low manipulation scores, the lowest scoring at 0.04, and slip past the detector entirely, direct evidence that a motivated, hands-on adversary can defeat current systems. These findings position the corpus as a diagnostic benchmark for studying not just whether authentication systems succeed, but where they fail and which failure modes matter for evidentiary use. 5. Ethical Considerations The corpus derives from publicly released research datasets and is redistributed in accordance with source licenses. No imagery was collected from human participants and no material originates from real legal proceedings. As the work involves secondary use of data with no participant interaction, it did not require research ethics board review. We draw only on the normal-activity subset of UCF-Crime; no individual appears in connection to a crime and no manipulation was designed to place an identifiable person at a real location or to attribute conduct to them. Manipulations were produced using capabilities already available through consumer interfaces so we judge the uplift from release to be small relative to the benefit to detection research. 6. Concluding Remarks This paper introduces the CIFAR Synthetic Evidence Corpus for evidentiary image authentication. The corpus enables diagnostic evaluation of contemporary AI manipulations across diverse photographic evidence, while our baselines show substantial weaknesses in current detectors, particularly for localized edits and authentic-image false positives. We release the corpus and benchmark infrastructure to support future research on reliable visual evidence authentication. Acknowledgements. This work was supported by the Safeguarding Courts from Synthetic AI Content Solution Network, funded through the Canadian AI Safety Institute (CAISI) Research Program at CIFAR. GenAI Usage Disclosure The generative models central to this work (GPT Image 2 (OpenAI), Gemini 3.1 with Nano Banana 2 (Google), and Firefly 5 (Adobe)) were used to construct the manipulated portion of the corpus, as described in Section 3. This use is the subject of the resource itself rather than an authoring aid, and all generated artifacts are released with the provenance metadata documented above. In preparing the manuscript, the authors used Claude to assist with improving the clarity of author-written prose and generating draft LaTeX for tables. All AI-assisted text was reviewed and edited by the authors, who take full responsibility for the content of the paper. No AI system is listed as an author. References Bray et al. (2023) S. D. Bray, S. D. Johnson, and B. Kleinberg Testing human ability to detect ‘deepfake’ images of human faces. Journal of Cybersecurity 9 (1), p. tyad011. External Links: ISSN 2057-2085, Link, Document Cited by: §1. Chandra et al. (2025) N. A. Chandra, R. Murtfeldt, L. Qiu, A. Karmakar, H. Lee, E. Tanumihardja, K. Farhat, B. Caffee, S. Paik, C. Lee, J. Choi, A. Kim, and O. Etzioni Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024. arXiv. Note: arXiv:2503.02857 [cs] version: 1 External Links: Link, Document Cited by: §2. Chesney and Citron (2018) R. Chesney and D. K. Citron Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security. SSRN Scholarly Paper, Social Science Research Network, Rochester, NY (en). External Links: Link, Document Cited by: §1, §2. Corvi et al. (2022) R. Corvi, D. Cozzolino, G. Zingarini, G. Poggi, K. Nagano, and L. Verdoliva On the detection of synthetic images generated by diffusion models. arXiv. Note: arXiv:2211.00680 [cs] External Links: Link, Document Cited by: §2. Dalal et al. (2025) A. Dalal, C. Gao, H. P. Grimm, M. Grossman, D. L. Jr, C. Pulice, V. S. Subrahmanian, and H. J. Tunheim Deepfakes in Court: How Judges Can Proactively Manage Alleged AI-Generated Material in National Security Cases. University of Chicago Legal Forum 2024 (1). External Links: ISSN 0892-5593, Link Cited by: §2. Delfino (2023) R. Delfino Deepfakes on Trial: A Call To Expand the Trial Judge’s Gatekeeping Role To Protect Legal Proceedings from Technological Fakery. UC Law Journal 74 (2), p. 293. External Links: ISSN 0017-8322<br />© Copyright University of California, College of the Law San Francisco, Link Cited by: §2. Dolhansky et al. (2020) B. Dolhansky, J. Bitton, B. Pflaum, J. Lu, R. Howes, M. Wang, and C. C. Ferrer The DeepFake Detection Challenge (DFDC) Dataset. arXiv. Note: arXiv:2006.07397 [cs] External Links: Link, Document Cited by: §2. Dong et al. (2013) J. Dong, W. Wang, and T. Tan CASIA Image Tampering Detection Evaluation Database. In 2013 IEEE China Summit and International Conference on Signal and Information Processing, p. 422–426. External Links: Link, Document Cited by: §2, §3. Ebrahimi et al. (2026) S. Ebrahimi, N. Jamali, B. Shirsalimian, K. McConvey, W. Zhang, J. Mahdavimoghaddam, M. Taranukhin, M. Grossman, V. Shwartz, Y. Deng, and E. Bagheri DetectZoo: A Unified Toolkit for AI-Generated Content Detection Across Text, Audio, and Image Modalities. arXiv. Note: arXiv:2606.04205 [cs.M] External Links: Link, Document Cited by: §4. Grimm et al. (2021) P. Grimm, M. Grossman, and G. Cormack Artificial Intelligence as Evidence. Northwestern Journal of Technology and Intellectual Property 19 (1), p. 9. External Links: ISSN 1549-8271, Link Cited by: §2. Grossman and Grimm (2025) M. Grossman and P. Grimm Judicial Approaches to Acknowledged and Unacknowledged AI-Generated Evidence. Science and Technology Law Review 26 (2) (en). External Links: ISSN 1938-0976, Link, Document Cited by: §1, §2. Guan et al. (2019) H. Guan, M. Kozak, E. Robertson, Y. Lee, A. Yates, A. Delgado, D. F. Zhou, T. N. Kheyrkhah, J. Smith, and J. G. Fiscus MFC Datasets: Large-Scale Benchmark Datasets for Media Forensic Challenge Evaluation. In IEEE Winter Conference on Applications of Computer Vision (WACV 2019), Waikola, HI (en). Note: Last Modified: 2019-12-31T06:12-05:00 External Links: Link Cited by: §2. He et al. (2021) Y. He, B. Gan, S. Chen, Y. Zhou, G. Yin, L. Song, L. Sheng, J. Shao, and Z. Liu ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis. arXiv. Note: arXiv:2103.05630 [cs] External Links: Link, Document Cited by: §2. Heaton et al. (2026) C. Heaton, S. Cleary, and M. Navin AI-generated evidence is a threat to public trust in the courts | National Center for State Courts. National Center for State Courts (en). External Links: Link Cited by: §1. Le et al. (2021) T. Le, H. H. Nguyen, J. Yamagishi, and I. Echizen OpenForensics: Large-Scale Challenging Dataset For Multi-Face Forgery Detection And Segmentation In-The-Wild. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, p. 10097–10107 (English). External Links: ISBN 978-1-6654-2812-5, Link, Document Cited by: §2. Li et al. (2020) Y. Li, X. Yang, P. Sun, H. Qi, and S. Lyu Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics. arXiv. Note: arXiv:1909.12962 [cs] External Links: Link, Document Cited by: §2. Mnookin (1998) J. Mnookin The Image of Truth: Photographic Evidence and the Power of Analogy. SSRN Scholarly Paper, Social Science Research Network, Rochester, NY (en). External Links: Link Cited by: §1. Oh et al. (2011) S. Oh, A. Hoogs, A. Perera, N. Cuntoor, C. Chen, J. T. Lee, S. Mukherjee, J. K. Aggarwal, H. Lee, L. Davis, E. Swears, X. Wang, Q. Ji, K. Reddy, M. Shah, C. Vondrick, H. Pirsiavash, D. Ramanan, J. Yuen, A. Torralba, B. Song, A. Fong, A. Roy-Chowdhury, and M. Desai A large-scale benchmark dataset for event recognition in surveillance video. In CVPR 2011, Colorado Springs, Colorado, p. 3153–3160. External Links: ISSN 1063-6919, Link, Document Cited by: §2, §3. Roca et al. (2025) T. Roca, A. C. Roman, J. T. Vega, M. Duarte, P. Wang, K. White, A. Misra, and J. L. Ferres How good are humans at detecting AI-generated images? Learnings from an experiment. arXiv. Note: arXiv:2507.18640 [cs] version: 1 External Links: Link, Document Cited by: §1. Rössler et al. (2019) A. Rössler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner FaceForensics++: Learning to Detect Manipulated Facial Images. arXiv. Note: arXiv:1901.08971 [cs] External Links: Link, Document Cited by: §2. Ryan and Garrett (2023) W. A. Ryan and A. Garrett Practical Lessons from the Attorney AI Missteps in Mata v. Avianca. Association of Corporate Counsel (en). External Links: Link Cited by: §1. Sultani et al. (2019) W. Sultani, C. Chen, and M. Shah Real-world Anomaly Detection in Surveillance Videos. arXiv. Note: arXiv:1801.04264 [cs] External Links: Link, Document Cited by: §2, §3. Yu et al. (2020) F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, and T. Darrell BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning. arXiv. Note: arXiv:1805.04687 [cs] External Links: Link, Document Cited by: §2, §3.