Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning

View on arXiv ← Back to list

Authors: Hao Tan, Jun Lan, Zichang Tan, Ajian Liu, Chuanbiao Song, Senyuan Shi, Huijia Zhu, Weiqiang Wang, Jun Wan, Zhen Lei

Published: 2025-08-28 17:53:05+00:00

Comment: Project: https://github.com/EricTan7/Veritas

AI Summary

The paper introduces HydraFake, a new dataset designed to simulate real-world deepfake challenges with hierarchical generalization testing, addressing discrepancies in existing benchmarks. Building upon this, they propose Veritas, a multi-modal large language model (MLLM) based deepfake detector that employs pattern-aware reasoning with planning and self-reflection. Veritas, trained via a two-stage pipeline, achieves significant generalization gains across diverse out-of-distribution scenarios and provides transparent, faithful detection outputs.

Abstract

Deepfake detection remains a formidable challenge due to the complex and evolving nature of fake content in real-world scenarios. However, existing academic benchmarks suffer from severe discrepancies from industrial practice, typically featuring homogeneous training sources and low-quality testing images, which hinder the practical deployments of current detectors. To mitigate this gap, we introduce HydraFake, a dataset that simulates real-world challenges with hierarchical generalization testing. Specifically, HydraFake involves diversified deepfake techniques and in-the-wild forgeries, along with rigorous training and evaluation protocol, covering unseen model architectures, emerging forgery techniques and novel data domains. Building on this resource, we propose Veritas, a multi-modal large language model (MLLM) based deepfake detector. Different from vanilla chain-of-thought (CoT), we introduce pattern-aware reasoning that involves critical reasoning patterns such as planning and self-reflection to emulate human forensic process. We further propose a two-stage training pipeline to seamlessly internalize such deepfake reasoning capacities into current MLLMs. Experiments on HydraFake dataset reveal that although previous detectors show great generalization on cross-model scenarios, they fall short on unseen forgeries and data domains. Our Veritas achieves significant gains across different OOD scenarios, and is capable of delivering transparent and faithful detection outputs.

Key findings

Experiments on HydraFake reveal that previous detectors generalize well on cross-model scenarios but struggle with unseen forgeries and data domains. Veritas achieves significant performance gains across various out-of-distribution (OOD) scenarios, outperforming state-of-the-art detectors by an average of 6.0% in accuracy. The pattern-aware reasoning and two-stage training pipeline are crucial for Veritas's improved generalization and its ability to deliver transparent, human-aligned detection outputs.

Approach

The authors introduce HydraFake, a new dataset simulating real-world deepfake challenges with hierarchical generalization testing. They propose Veritas, an MLLM-based deepfake detector that utilizes pattern-aware reasoning, incorporating critical thinking patterns like planning and self-reflection to emulate human forensic processes. This is achieved through a two-stage training pipeline involving pattern-guided cold-start (SFT with Mixed Preference Optimization) and pattern-aware exploration (Pattern-Aware Group Relative Policy Optimization).

Datasets

HydraFake-100K (newly introduced), LFW, CelebA, FaceForensics++, FFIW, FFHQ, VFHQ, UADFV, CelebAHQ, DF40, WILD, seeprettyface, Talking-HeadBench, and self-constructed deepfakes from advanced generators and commercial apps like GPT-4o, Dreamina, and Hailuo AI.

Model(s)

InternVL3-8B (as the base MLLM for Veritas)

Author countries

China

← Previous