← Back to Blog

Research RadarDeepfake DetectionarXivOctober 2026

Monthly arXiv Radar

October 2026 Deepfake Detection Papers: Broad Forensics, Long Videos, and Partial Faces

October's detection research broadens the threat model in three directions. Venus-DefakerOne2 unifies media and attack forms, AV-FABLE asks systems to locate short audio and visual edits inside long open-scene video, and RLDF tests whether a detector notices when only a small facial region is changed.

What This Month Signals

Venus-DefakerOne2 reports one model across six forensic domains, including 96.7% balanced accuracy on deepfake faces and 96.0% on AI-generated video, though its scale and internal data make independent validation essential. AV-FABLE adds 998 hours of long-form twin videos with separate audio and visual timelines; zero-shot language outputs remain near chance while supervised readouts of frozen features reach 91.30 mAP. RLDF shows that limiting a swap to a facial sub-region drops existing detectors by 10.5 AUC points on average. The shared procurement lesson is to test threat coverage, temporal localization, modality attribution, and manipulated-area size rather than accepting one aggregate accuracy.

Paper 012026-10-27cs.CV

Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage

Authors & Institutions

Chenfan Qu

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Yachao Liang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Mingqi Fang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Jiangwei Xie

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Zhenming Wang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Baoyue Zhang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Changjiang Jiang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Zheng Chang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Fanli Jin

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Xuekang Zhu

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Kaiwen Feng

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Junchi Li

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Weihong Zhang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Haoyue Wang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Chenfeng Zhang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Yinan Xia

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Fangming Cui

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Yongshuo Zhang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Dong Fan

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Jingjing Liu

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Tianyi Zhang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Luobin Chen

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Ji-Zhe Zhou

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Sheng Li

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Wei Wang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Bin Li

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Ran He

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Xinpeng Zhang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Linna Zhou

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Jian Liu

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

Weiqiang Wang

GuangJian Team, Ant Group

Technical Committee on Digital Media Forensics and Security, CSIG

What Problem It Solves

Venus-DefakerOne2 aims to provide one binary forensic foundation model across digital and physical forgery in images and videos. The challenge is to add domains and difficult samples without catastrophic forgetting or forcing one prediction interface onto every media type.

Key Result

The paper reports balanced accuracy of 96.7% for deepfake faces, 93.3% for AI-generated images, 87.6% for tampered documents, 89.6% for manipulated natural images, 93.1% for physical spoofs, and 96.0% for AI-generated videos. On the two highlighted broad domains, physical forgery improves from DefakerOne's 60.8% to 93.1%, and AI-video detection from 72.5% to 96.0%.

Abstract

Real-world forgery is inherently heterogeneous: it manifests not only as digital manipulation but also as physical spoofing, and is carried not only by images but also by videos. Despite substantial progress, a unified anti-forgery foundation model covering such a broad threat spectrum remains absent. In this paper, we present Venus-DefakerOne2, which substantially extends the forensic capability boundary along both the forgery-form and media-type dimensions. Trained on tens of millions of open-source and in-house samples, it incorporates two key designs. Tide-Style-Tuning alternates LoRA and full-parameter fine-tuning to balance computational efficiency, generalization, and hard-sample fitting while mitigating catastrophic forgetting. Entropy-based Test Adaptation determines the optimal prediction modality by analyzing prediction entropy. Venus-DefakerOne2 achieves new state-of-the-art performance across a comprehensive suite of benchmarks, including DeepFake faces (96.7%), AI-generated images (93.3%), tampered documents (87.6%), manipulated natural images (89.6%), physical spoofs (93.1%), and AI-generated videos (96.0%). It significantly outperforms both specialized expert models and its predecessor, DefakerOne, establishing itself as a powerful and versatile foundation model for binary-decision forensic tasks.

Research Starting Point

Operational forgery screening spans deepfake faces, generated images and videos, edited documents, natural-image tampering, and physical presentation attacks. Separate specialist models increase integration cost and can leave inconsistent blind spots across media and attack forms.

Method

The model is trained on tens of millions of open-source and internal samples. Tide-Style-Tuning alternates parameter-efficient LoRA phases with full-parameter fine-tuning to preserve broad generalization while fitting hard cases. Entropy-based test adaptation compares available prediction modalities and selects the lower-uncertainty route; additional components support video aggregation and pixel-level localization.

Paper Summary

A broad foundation detector can simplify a multi-media review stack, but the headline average spans many benchmarks and includes internal training data. Enterprise evaluation still needs attack-specific thresholds, calibration, latency, and testing on unseen generators and capture devices.

Paper 022026-10-24cs.CV

AV-FABLE: Audio-Visual Manipulation Forensics in Long-Form Videos

Authors & Institutions

Yuesheng Huang

Tsinghua University

Hoi Fan Au

Tsinghua University

Jiaqi Hou

Tsinghua University

Ruicheng Zhang

Tsinghua University

Jiarui He

Tsinghua University

Haocheng Yang

Tsinghua University

Bin Hu

Tsinghua University

Kewei Zhang

Peking University

Weijian Deng

Tsinghua University

Jun Zhou

Tsinghua University

Xiu Li

Tsinghua University

What Problem It Solves

AV-FABLE defines separate audio and visual localization timelines for sparse manipulations in open-scene, minute-scale video. It tests detection, temporal localization, modality attribution, and evidence generation in domain and under source, generator, route, and external-dataset shift.

Key Result

The benchmark contains 16,650 pristine-manipulated twin pairs, 33,300 videos, and 998.1 hours; the median video is 88 seconds and manipulation covers 9.45% of it. Zero-shot audio-visual language models reach only 51-52 AUROC and at most 0.30 mAP, although linear probes recover 90.8 audio and 80.8 visual AUROC. NFSR reaches 91.30 audio-visual mAP and 96.04 detection AUROC in domain and leads all three out-of-distribution settings.

Abstract

Localized AI manipulations in long-form videos often affect brief events within otherwise authentic content, making video-level classification insufficient for forensic analysis. Existing audio-visual deepfake benchmarks largely focus on short talking-head clips, while long-form temporal forensics remains predominantly visual. We introduce AV-FABLE, a dual-timeline benchmark for sparse manipulations in open-scene long-form audio-visual videos. AV-FABLE contains 16,650 pristine--manipulated twin pairs (998 hours) created with 27 editing configurations, with separate audio and visual timelines that enable modality-specific temporal localization and expose errors hidden by joint-timeline evaluation. On AV-FABLE, zero-shot prompting of audio-visual large language models (AV-LLMs) yields near-chance detection and poor temporal localization. In contrast, supervised readouts of their frozen perception features recover strong forensic signals, suggesting manipulation evidence is encoded in these representations but not effectively exposed through language outputs. Motivated by this observation, we propose NFSR(Native Forensics via Separated Readouts), which reads frozen AV-LLM features with separate audio and visual temporal decoders. NFSR achieves the best audio-visual localization performance in-domain and across all three out-of-distribution settings, including transfer to held-out videos from an external long-video forensic benchmark. AV-FABLE, code, and model weights will be publicly released.

Research Starting Point

A few forged seconds inside a mostly authentic long video can change its meaning, and audio and visual edits may occur at different times. A single video label or shared timeline cannot tell an investigator which modality was altered and may reward a model for finding the right moment on the wrong stream.

Method

SEE-Forge selects natural events, plans a localized edit, routes it to one of 27 audio or visual editing configurations, reinserts it into context, and pairs every manipulated output with a matched pristine export. NFSR then reads frozen audio-visual language-model features with separate temporal decoders, combining intermediate and interface representations and conditioning local evidence on the host video.

Paper Summary

AV-FABLE is a strong test for review tools that must point to the forged interval and affected stream, not just flag a file. It also warns that useful forensic evidence inside a foundation model may not be accessible through zero-shot prompting.

Paper 032026-10-15cs.CV

RLDF: A Region-Level Deepfake Dataset and Benchmark for Partial-Face Manipulation Detection

Authors & Institutions

Shukesh Reddy

Machine Intelligence Group, BITS Pilani, Hyderabad Campus

Abhijit Das

Machine Intelligence Group, BITS Pilani, Hyderabad Campus

What Problem It Solves

RLDF provides region labels and spatial ground truth for partial-face swaps and audits whether established detectors transfer from full-face benchmarks. It also supplies a localization-aware baseline for learning from the small forged area rather than averaging it away.

Key Result

Across 12 detectors, region-level testing causes an average 10.5 percentage-point AUC drop and 10-point accuracy drop relative to conventional evaluation. FatFormer falls to 0.745 AUC and DFD-FCG to 0.806 on RLDF. The mask-guided baseline establishes that explicit spatial supervision can recover useful signal, but the paper positions it as a starting point rather than a solved detector.

Abstract

Existing deepfake detection models are built upon manipu- lated images that are tampered with either exclusively fully swapped or fully synthesised faces, which leads us to an open question: Can off-the-shelf deepfake models detect manipulation effectively while tampering is confined to a local facial sub-region? To fill this gap, we introduce RLDF (Region-Level Deepfake), the first dataset for region-level deepfake evaluation. Employing the proposed region-aware partial swap synthesis model, we generate over 2M par- tial face-swap images by using eight face-swapping meth- ods. Altogether, we considered seven face regions, among which there are four discrete regions (eye, nose, mouth and eyebrow) and four composite facial regions (eye and nose; eye, nose and mouth; eye, nose, mouth and eye- brow; and full face). In addition, we define three evalu- ation protocols and benchmark 12 state-of-the-art detectors across the five popular standard cross-dataset evaluations, DF40, and the proposed RLDF, providing a comprehen- sive audit of where and why existing detectors underper- formed. Our evaluation reveals that existing detectors fail to generalize consistently across partial facial regions, re- sulting in an overall drop of 10.5% AUC and 10% accu- racy under region-level evaluation. Further, we propose a mask-guided cross-attention model as a baseline approach that can effectively detect region-level manipulation. The proposed dataset, code, and checkpoints are available at https://github.com/RLDF-Dataset/RLDF.

Research Starting Point

Most deepfake benchmarks replace or synthesize an entire face, so detectors can learn broad global artifacts. Modern editing tools can alter only the eyes, mouth, nose, or another small region, leaving authentic pixels to dominate the image and making a confident global classifier much less reliable.

Method

The dataset applies eight face-swapping methods to 1,000 source-target video pairs, then uses face parsing, dilation, and soft blending to composite eight individual or composite facial regions back into authentic target frames. This produces 2,048,000 fake images with identity-disjoint train, validation, and test splits. A CLIP ViT-L/14 baseline predicts a patch forgery mask and uses it to guide cross-attention before the final real-or-fake decision.

Paper Summary

Partial-face manipulation should be a separate acceptance test for any deepfake product. RLDF demonstrates that strong full-face scores do not guarantee sensitivity to a small edited region and provides the masks needed to train and inspect localized decisions.