← Back to Blog
Research RadarDeepfake DetectionVideo ForensicsarXivAugust 2026

Monthly arXiv Radar

August 2026 Deepfake Detection Papers: Image-to-Video Transfer, Multi-Agent Forensics, and Fairness

The August papers treat deepfake detection as a deployment system rather than a single AUC number. One study shows that image detectors can outperform dedicated video models with frame tuning and learned aggregation; another separates texture, lighting, motion, and physics analysis across agents; the third targets demographic shortcuts inside spatial and fine-tuning residual features.

What This Month Signals

FakeI2V-Bench finds that video-frame tuning plus statistical aggregation can turn mature image detectors into stronger and sometimes lighter video systems, but sparse edits remain difficult. FaceVid-Forensics-100K uses four independent forensic agents and a judge to improve unseen-generator detection while returning explanations. FairReL shows that generalization and fairness are separate objectives: a detector can transfer well overall and still concentrate false positives on minority subgroups. A credible production evaluation therefore needs generator diversity, temporal sparsity tests, reviewable evidence, and fixed-threshold subgroup metrics.

Paper 012026-08-04cs.CR

FakeI2V-Bench: Benchmarking the Applicability of Image-level Deepfake Detectors for Deepfake Video Detection

arXivPDF

Authors & Institutions

Pei Li

School of Cyber Science and Technology, Shandong University, Qingdao, China

Sihan Chen

School of Cyber Science and Technology, Shandong University, Qingdao, China

Delong Ran

Institute for Network Sciences and Cyberspace, BNRist, Tsinghua University, Beijing, China

Tianshuo Cong

School of Cryptologic Science and Engineering, Shandong University, Jinan, China

What Problem It Solves

FakeI2V-Bench supplies a unified test of image- and video-level detection and introduces a reusable bridge that turns frame scores into a video decision. It quantifies accuracy, cross-generator behavior, latency, model size, and performance when only 20% of a clip is manipulated.

Key Result

With the best naive aggregation, the strongest image detector reaches 80.16% AUC, already edging the top video model FTCN at 79.99%. After IV-Bridge, eleven of twelve image detectors beat FTCN; RINE-IV leads at 93.80% AUC and 97.63% AP, while the 4.34-million-parameter Patch-IV reaches 89.26% AUC and 96.12% AP. Across fourteen generators, enhanced image models average 95.51% AUC versus 91.01% for the best video detector, although Sora remains difficult at 72.76% mean AUC. With only 20% forged frames, every model drops and Patch-IV leads at 77.15% AUC.

Abstract

Recent advances in video generation models have significantly intensified the deepfake threat, yet the current deepfake video detection benchmarks remain underdeveloped. In particular, the effectiveness of image-level detectors in the video domain has not been systematically assessed. To fill this gap, we present FakeI2V-Bench, a benchmark for evaluating state-of-the-art video-level deepfake detectors in challenging scenarios, with a particular focus on systematically assessing the performance of image-level deepfake detectors in the video domain. FakeI2V-Bench comprises 97,548 videos, containing content generated by the latest powerful generation models and covering a broader range of categories. Using this dataset, we conduct a systematic evaluation of eight video-level detectors and twelve representative image-level detectors. Experimental results show that the best-performing image-level detector achieves an 80.16% AUC, slightly outperforming the strongest video-level detector (i.e., 79.99% AUC). Going beyond benchmarking, we present IV-Bridge, a general framework that enhances the applicability of image-level deepfake detectors to videos. IV-Bridge employs a random forest model with statistical features to aggregate frame-level predictions, allowing eleven image-level detectors to surpass state-of-the-art video-level approaches, with the best-performing variant achieving a 93.80% AUC. Overall, FakeI2V-Bench establishes a rigorous benchmark for deepfake video detection and introduces a novel pathway for extending image-level detectors to the video domain, offering new insights and directions for future research. Code and data are available at https://github.com/CryptoAILab/FakeI2V-Bench.

Research Starting Point

Image deepfake detectors benefit from mature training pipelines, but video benchmarks usually compare only video-specific architectures and apply frame models with simplistic averaging. That leaves teams unsure whether temporal networks are necessary, how recent video generators change the ranking, and whether an existing image detector can be adapted economically to video. Sparse manipulations, where only a fraction of frames are fake, make naive aggregation especially risky.

Method

The benchmark combines 97,548 videos from four facial and general-generation datasets and evaluates eight video detectors plus twelve representative image detectors. The authors first compare fixed score aggregations such as mean, maximum, and minimum. IV-Bridge then performs video-frame fine-tuning on non-overlapping FF++ and GenVideo frames and feeds statistical descriptors of frame-level scores to a random forest with as many as 800 trees and depth eight. Results are broken down by fourteen generation models, facial versus general content, deployment cost, and short-term forgery.

Paper Summary

An existing image detector can become a competitive video service if it is tuned on real video frames and paired with learned temporal score aggregation. The bridge is attractive for lightweight deployments, but sparse edits remain a clear failure mode and the largest accuracy winner is not the cheapest model.

Paper 022026-08-07cs.CV

Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection

arXivPDF

Authors & Institutions

Xuechao Zou

Beijing Jiaotong University, China

Shun Zhang

Beijing Jiaotong University, China

Kai Li

Tsinghua University, China

Yi Zhou

Beijing Jiaotong University, China

Xinyu Sun

Beijing Jiaotong University, China

Yuhui Chen

Ant Group, China

Zhe Wu

Tsinghua University, China

Congyan Lang

Beijing Jiaotong University, China

Junliang Xing

Tsinghua University, China

What Problem It Solves

The paper contributes both a current, explanation-rich benchmark and a collaborative reasoning system designed for unseen generators. It tests whether independent specialist reports improve generalization beyond longer chain-of-thought prompts or repeatedly showing the same frames to one model.

Key Result

On the out-of-domain test, the full system with video-aware judging reaches 69.87% accuracy, 81.82% recall, and 53.28 F1. The strongest listed closed model, Gemini 2.5 Pro, records 63.78%, 75.29%, and 47.45%, while the best small vision baseline reaches 64.28% accuracy and 45.20 F1. A text-only judge over agent reports still ranks second at 67.41% accuracy and 51.01 F1. Joint SFT improves the video-aware system from 42.29% to 67.03% accuracy, and GRPO raises it further; combining all four forensic agents gives the best F1 under both judge settings.

Abstract

The malicious use of generative artificial intelligence to create highly realistic deepfake videos raises serious ethical concerns and poses substantial challenges to AI safety. However, existing deepfake video benchmarks provide limited coverage of recent synthesis methods and generally lack reliable fine-grained textual annotations. Meanwhile, conventional detectors and multimodal large language models (MLLMs), whether operating as a single model or relying on a single analytical perspective, often fail to capture subtle forgery artifacts, limiting their generalization to emerging AI-generated methods. To address these limitations, we introduce FaceVid-Forensics-100K, a large-scale deepfake video dataset comprising 100,000 videos and spanning 33 synthesis methods across face swapping, face reenactment, and entire-face synthesis, including recent generators such as Seedance 2.0. The dataset provides fine-grained textual annotations of visual observations and verdict-consistent forensic explanations, automatically synthesized through a multi-model aggregation and conflict-resolution pipeline powered by advanced MLLMs. Building on this benchmark, we propose a multi-agent forensic reasoning framework that employs four specialized domain-expert agents to independently analyze forgery cues from four perspectives: texture, lighting, motion, and physics. A judge agent then reconciles their reports to produce a final prediction together with an explanation. Extensive evaluations on out-of-domain test sets show that, despite being composed entirely of small open-source MLLMs, our framework outperforms all methods including closed-source GPT and Gemini models and ranks first across all reported metrics on this benchmark. The project page is available at https://xavierjiezou.github.io/ARGUS/.

Research Starting Point

Modern face video generators maintain identity and motion well enough that a single classifier or general-purpose MLLM can miss subtle, generator-specific evidence. Existing datasets often lag new synthesis systems and provide only binary labels, making it hard to train or audit explanations. For high-risk review, a detector should inspect independent forensic dimensions and reconcile conflicting observations rather than produce an unsupported verdict.

Method

FaceVid-Forensics-100K contains 100,000 videos from 33 face-swapping, reenactment, and whole-face synthesis methods, including Seedance 2.0. Fine-grained observations and verdict-consistent explanations are created with multi-model aggregation and conflict resolution. Four small open-source MLLM agents separately inspect texture, lighting, motion, and physical consistency; a judge receives their reports, optionally samples the video, and emits a verdict plus explanation. The agents and judge are trained with supervised fine-tuning and then group relative policy optimization, with out-of-domain generators held out for evaluation.

Paper Summary

Separating forensic observation from verdict formation improves both auditability and unseen-generator performance. The architecture is compelling for analyst-assist workflows, but its multi-model inference cost and automatically generated explanation labels should be measured against a simpler detector plus human review before production adoption.

Paper 032026-08-28cs.CV

FairReL: Deepfake Detection using Fairness-Aware Representation Learning

arXivPDF

Authors & Institutions

Xiaoman Lu

Department of Computer Science, University of Warwick, Coventry, United Kingdom

Jiaqi Li

Department of Computer Science, University of Warwick, Coventry, United Kingdom

Shuntian Zheng

Department of Computer Science, University of Warwick, Coventry, United Kingdom

Huiping Chen

School of Computer Science, University of Birmingham, Birmingham, United Kingdom

Yu Guan

Department of Computer Science, University of Warwick, Coventry, United Kingdom

What Problem It Solves

FairReL isolates two subgroup-sensitive components - multi-scale spatial frequencies and the residual representation introduced during fine-tuning - and applies different fairness controls to each. It evaluates false-positive disparity at a fixed operating threshold so improvements cannot be manufactured by simply rejecting fewer samples overall.

Key Result

Against the strongest fairness-oriented baseline, the authors report a 3.9-point improvement in unseen-dataset AUC and a 10.2% reduction in subgroup FPR disparity. FairReL reaches AUCs of 80.64, 86.13, and 66.34 on Celeb-DF, DFD, and DFDC, while posting the lowest equal-FPR disparity on all three (10.12, 13.55, and 38.36). On DFDC its group-averaged AUC is 67.03 and its best-to-worst subgroup AUC gap is 1.09 points. It slightly trails the accuracy-focused GenD in overall AUC on two sets but distributes performance more evenly, especially for female and Black subgroups.

Abstract

Although recent deepfake detectors achieve high overall accuracy, their errors remain unevenly distributed across demographic subgroups, with real faces from certain groups more often misclassified as fake. Existing fairness-aware detectors typically regularise the entire feature representation, without identifying or controlling the specific components that drive unfair predictions. Such coarse intervention can over-suppress useful forgery cues while leaving demographic structure in component-specific subspaces. To address this, we identify two subgroup-sensitive components: multi-scale spatial features, which encode local facial and forgery patterns, and fine-tuning-induced residual features, which adapt the backbone to the unfair training distribution. We propose FairReL, a fairness-aware representation-learning framework that targets both components with dedicated demographic supervision. FairReL uses an SVD-decomposed foundation-model backbone to isolate the fine-tuning-induced residual representation, and introduces two complementary losses. Group-Conditional Wavelet Decorrelation (GCWD) suppresses subgroup-imbalanced structure across spatial wavelet sub-bands, while Subspace-Localised Mean Alignment (SLMA) aligns subgroup means within each real/fake class in the residual representation. Experiments on FF++, Celeb-DF, DFD and DFDC show that, against the state-of-the-art fairness-aware detector, FairReL improves unseen-dataset AUC by 3.9% while reducing subgroup FPR disparity by 10.2%. Code is available at https://github.com/xiaoman89/FairReL .

Research Starting Point

A detector can achieve strong global AUC while falsely accusing members of some demographic subgroups much more often than others. Existing fairness losses often regularize the final prediction or the entire embedding, which can leave localized demographic shortcuts intact or erase useful forgery evidence. The challenge is to target where subgroup information enters the detector and preserve both fairness and cross-domain accuracy.

Method

The method decomposes every self-attention projection in CLIP ViT-L/14 with SVD and learns a rank-one residual while retaining the foundation representation. Group-Conditional Wavelet Decorrelation suppresses subgroup-imbalanced structure separately across spatial wavelet bands. Subspace-Localized Mean Alignment aligns demographic means within each real/fake class only in the adaptation residual, avoiding indiscriminate full-feature smoothing. Models train on FaceForensics++ and transfer to Celeb-DF, DFD, and DFDC, using six intersectional gender/ethnicity groups and FPR metrics at a source-validation threshold fixed to 90% TPR.

Paper Summary

Deepfake evaluation should report subgroup false positives at a fixed operating point, not only aggregate AUC. FairReL shows that localized debiasing can retain competitive cross-domain detection, though the residual DFDC disparity remains large enough that production thresholds and escalation paths still need subgroup-aware validation.