Research RadarDeepfake DetectionFace ForensicsarXivJuly 2026

Monthly arXiv Radar

July 2026 Deepfake Detection Papers: Reconstruction Evidence, Facial Physiology, and Compact Replay

July's deepfake papers address three weaknesses that appear after a detector leaves a static benchmark: explanations can be ungrounded, talking-face generators can evade artifact classifiers, and continual updates can forget earlier manipulation families. The selected methods introduce an explicit reconstruction residual, a physiological video channel, and an information-dense replay buffer.

What This Month Signals

LaP-Forensics grounds localization and structured explanations in a diffusion reconstruction residual, but correctly avoids claiming that its free-form text is fact-verified. The rPPG study shows that talking-face detection difficulty varies dramatically by generator even when the classifier and protocol remain fixed. InfoDense addresses the operational problem of updating a detector without storing full historical face images. Across all three, the strongest signal is not a universal artifact: reliability comes from combining complementary evidence, isolating evaluation protocols, and preserving the cues that matter when distributions evolve.

Paper 012026-07-28cs.CV

LaP-Forensics: Latent-Pixel Consistency Guided Multimodal Reasoning for Deepfake Detection

arXiv PDF

Authors & Institutions

Can Wang

The Hong Kong Polytechnic University, Hong Kong, China

Yuhao Wang

University College London, United Kingdom

Yushe Cao

Tsinghua University, China

Canran Xiao

Sun Yat-sen University, China

Fei Shen

National University of Singapore, Singapore

What Problem It Solves

LaP-Forensics addresses cross-generator image detection and pixel-level artifact localization with an explicit compatibility reference. It aims to make structured where-what-why reasoning consult that evidence, while acknowledging that reference use does not by itself prove every sentence in a free-form explanation.

Key Result

On UniversalFakeDetect, the image head reaches 97.23% accuracy on GANs, 71.30% on Deepfakes, 92.18% on Diffusion, and 98.98% on CRN. Under the official SynthScars protocol, human-artifact localization reaches 61.40 mIoU and 66.00 F1. In the controlled split, removing the residual lowers mIoU from 72.19 to 61.83 and F1 from 63.62 to 41.71; replacing the residual with zero or a donor map also materially shifts the output.

Abstract

Recent generative models can produce images with few obvious visual artifacts, weakening detectors and explanations that rely only on surface appearance. We present LaP-Forensics, a multimodal framework that augments RGB semantics with reconstruction-based forensic evidence. A frozen Stable Diffusion DDIM inversion-reconstruction model provides a fixed reconstruction reference, and its residual map measures local compatibility with that reference. Independent projectors encode the RGB image and residual map before a structured Where-What-Why model predicts a textual analysis and an artifact mask.Supervised fine-tuning is followed by Group Relative Policy Optimization (GRPO), whose reward combines mask overlap with output-structure and evidence-reference terms. These text-side terms encourage the model to refer to the consistency map but do not constitute a verifier of free-form textual truth. A separate image-level head fuses RGB and DDIM-residual class features. Experiments show cross-generator detection on UniversalFakeDetect and competitive artifact localization on the official SynthScars benchmark. Controlled cue-construction, inversion-horizon, component, reward-term, and counterfactual analyses support the utility of the residual stream under the evaluated settings, while free-form textual faithfulness and reliability under post-processing remain open limitations.

Research Starting Point

As generated images lose obvious surface defects, an RGB-only detector or vision-language explanation can sound persuasive without pointing to stable forensic evidence. For enterprise review, a useful system should separate the detection score, spatial localization, and textual account while showing what additional signal actually changes the result.

Method

A frozen Stable Diffusion DDIM inversion-reconstruction process produces a residual map between the input and reconstruction. Separate projectors encode RGB semantics and residual evidence; a multimodal branch generates structured text and a mask, while a separate image head fuses class features for detection. Supervised fine-tuning is followed by GRPO rewards for mask overlap, output format, and evidence references, with official benchmark tests kept separate from a smaller controlled ablation split.

Paper Summary

The reconstruction residual is a useful second forensic channel and improves localization measurably, but it is not a calibrated manipulation map or a source-attribution signal. Teams should value the clean separation of detection, localization, and explanation, then independently test post-processing robustness and textual faithfulness before using the generated rationale in review workflows.

Paper 022026-07-23cs.LG

Physiological Signals as a Forensic Modality for Talking-Face Deepfake Detection

arXiv PDF

Authors & Institutions

Othmane Harraq

Temple University, Philadelphia, Pennsylvania, USA

Tamer Aldwairi

Temple University, Philadelphia, Pennsylvania, USA

What Problem It Solves

The paper asks whether remote photoplethysmography is more informative for talking-face forgeries and whether performance remains credible when every test identity is absent from training. It also measures detector difficulty separately for seven generation methods instead of hiding that variation inside one aggregate score.

Key Result

The 165K-parameter 1D ResNet reaches AUC 0.806 +/- 0.003 and EER 27.8%, versus AUC 0.622 for the reproduced prior rPPG detector and 0.830 for the cited general-purpose Effort model. Per-generator AUC ranges from 0.985 for Real3DPortrait to 0.690 for IP-LAP, a stable 0.295 spread that is much larger than seed variance.

Abstract

Talking-face (TF) deepfake generation synthesizes photore- alistic facial video from a static source image and an au- dio signal, producing forgeries that current image-based detectors consistently fail to identify. Unlike face-swap ma- nipulation, TF synthesis has no underlying real video from which to inherit physiological characteristics, making re- mote photoplethysmography (rPPG) a uniquely motivated detection modality for this forgery category. We propose a detection framework that extracts per-video rPPG wave- forms via RhythmFormer and trains a suite of lightweight classifiers to distinguish real from synthesized physiologi- cal signals. Evaluated on the TF subset of Celeb-DF++ un- der a strict subject-independent protocol, where test identi- ties are completely separated from training identities, our 1D ResNet achieves an AUC of 0.806 and EER of 27.8%, placing it within 2.4 points of the best published general- purpose detector (Effort, ICML 2025) while operating ex- clusively on the physiological channel. We document a con- trolled reproduction study of DeepFakesON-Phys, the rep- resentative prior rPPG detector, demonstrating degrada- tion from AUC 0.999 on legacy face-swap data to 0.622 on the TF subset of Celeb-DF++. We further show that detec- tion difficulty is strongly method-dependent: AUC ranges from 0.985 (Real3DPortrait) to 0.690 (IP-LAP) across the seven TF generators, with the ranking remaining perfectly stable across all evaluation protocols. This spread reflects an interpretable physiological property of each generator rather than evaluation noise, and constitutes the primary theoretical contribution of the work.

Research Starting Point

Talking-face generators synthesize an entire video from a still identity and audio, so they may not preserve a real temporal blood-volume signal. Existing rPPG detectors were mainly validated on older face swaps, where underlying source-video physiology can survive and where identity leakage across splits can inflate results.

Method

RhythmFormer extracts a waveform from each face video, after which lightweight one-dimensional ResNet and transformer classifiers distinguish real from synthetic physiology. The Celeb-DF++ talking-face subset is split by celebrity identity, with 18 identities held out for evaluation. The authors reproduce DeepFakesON-Phys under the same setting and run five-seed aggregate and per-generator analyses.

Paper Summary

Facial physiology is a complementary, interpretable channel rather than a standalone answer. It comes close to a larger visual detector with very few parameters, but some generators preserve pseudo-periodic signals well enough to approach chance; production systems should fuse rPPG with spatial, frequency, and audio-visual evidence and validate across identities and generators.

Paper 032026-07-18cs.CV

InfoDense: Density-Aware Regional Decisive Replay for Memory-Efficient Incremental Face Forgery Detection

arXiv PDF

Authors & Institutions

Jikang Cheng

Peking University, China

Hao Shen

Huazhong Agricultural University, China

Xueyi Zhang

National University of Singapore, Singapore

Guangcheng Wang

Nantong University, China

Zhongyuan Wang

Wuhan University, China

Renye Yan

Peking University, China

Baojin Huang

Huazhong Agricultural University, China

What Problem It Solves

InfoDense reframes replay as an information-density problem: retain compact regions that carry manipulation evidence, select a diverse set of those regions, and combine them with current-task samples without introducing the domain bias associated with a buffer of old full images.

Key Result

Under the main protocol, InfoDense reports 90.42% average AUC across incremental datasets with a 4.52-point performance drop, and 90.03% average cross-dataset AUC. The paper reports comparable performance after an 80% storage reduction and up to ten times the replay diversity at equal storage. Stored fragments also reduce identity retrieval recall to below 0.02% in the reported privacy analysis.

Abstract

The rapid evolution of face forgery techniques has introduced an increasing variety of manipulations. Incremental Face Forgery Detection (IFFD), which incrementally adds new forgery data to fine-tune previously trained models, has emerged as a promising approach to handle evolving forgery threats. However, conventional replay-based IFFD methods suffer from catastrophic forgetting. Storing full historical images under limited memory often either fails to preserve subtle forgery cues or introduces domain bias, reducing the model's ability to learn intrinsic and transferable manipulation characteristics. In this paper, we propose a Density-Aware Regional Decisive replay strategy, termed InfoDense, to address these challenges. InfoDense prioritizes artifact-dense and forgery-critical regions, significantly reducing storage requirements while maintaining high-fidelity forgery evidence. We first introduce InfoDense Cut to localize decisive patches using CLIP-based embeddings. Then, InfoDense Select ranks candidate segments by combining latent-space representativeness and decisive patch counts, ensuring both diversity and information density in the replay buffer. Finally, InfoDense Fuse reconstructs unbiased training inputs by adaptively merging stored segments with current-task samples, enhancing knowledge retention and generalization. Extensive experiments on challenging incremental deepfake benchmarks demonstrate that InfoDense effectively mitigates catastrophic forgetting while improving cross-domain generalization.

Research Starting Point

Deepfake detectors need repeated updates as new manipulation families arrive, but replaying full historical face images consumes memory, retains identifiable content, and still may spend most of the buffer on background or redundant facial structure. Under a fixed storage budget, a detector must preserve subtle forgery cues without forgetting earlier tasks.

Method

InfoDense Cut uses CLIP patch embeddings and forgery-related text to score decisive regions. InfoDense Select ranks candidate segments by both latent-space representativeness and the number of decisive patches. InfoDense Fuse reconstructs training inputs by combining stored historical segments with current samples, increasing the number of source images represented under the same memory budget. Evaluation follows incremental forgery sequences and separate cross-dataset tests.

Paper Summary

Region-level replay is attractive for a detector that must learn continuously under storage or privacy constraints. Its deployment value depends on whether the selected patches remain informative after real compression, policy changes, and adversarial post-processing, but it provides a concrete alternative to retaining complete historical faces.