Deepfake Detection

Self-consistency and language guidance for generalizable deepfake detection.

Two complementary approaches to detecting synthetic faces, designed for the deployment problem rather than the leaderboard problem.

Pair-wise self-consistency learning (Zhao et al., 2021; Zhao et al., 2020) (ICCV 2021 oral) trains a model to spot the source-feature inconsistency that GenAI generators leave behind, paired with an inconsistency image generator that synthesizes richly annotated training data. The hypothesis worth testing: even when a generator’s output is photorealistic to a human, the source features within the forged image preserve a fingerprint detectable by ConvNets.

AuthGuard (Shen et al., 2026) (WACV 2026) extends the same instinct with a language-guided expert encoder that reasons about logical and perceptual anomalies the way a person would, on top of the statistical artifacts. AUC gains of 6.15% on DFDC and 16.68% on DF40, with a 24.69% improvement on the DDVQA reasoning benchmark.

These ideas shaped the omni-modal trust work and the production model-diagnosis program that addresses the failure modes uncovered by deepfake detection at production scale.

References

2026

  1. AuthGuard: Generalizable Deepfake Detection via Language Guidance
    Guangyu Shen, Zhihua Li, Xiang Xu, and 6 more authors
    In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2026

2021

  1. Learning Self-Consistency for Deepfake Detection
    Tianchen Zhao, Xiang Xu, Mingze Xu, and 3 more authors
    In IEEE/CVF International Conference on Computer Vision (ICCV), 2021
    Oral

2020

  1. arXiv
    Learning to Recognize Patch-Wise Consistency for Deepfake Detection
    Tianchen Zhao, Xiang Xu, Mingze Xu, and 3 more authors
    arXiv preprint arXiv:2012.09311, 2020