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

AI Summary

This paper introduces HydraFake, a new deepfake detection dataset designed to better reflect real-world challenges, and Veritas, a multi-modal large language model (MLLM)-based deepfake detector that uses pattern-aware reasoning to improve generalization to unseen forgeries and data domains.

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

Veritas significantly outperforms state-of-the-art deepfake detectors on cross-forgery and cross-domain scenarios of the HydraFake dataset. The two-stage training pipeline, particularly the pattern-aware reasoning, is crucial for this improved generalization. The model also demonstrates robustness to image compression and blur.

Approach

Veritas uses a two-stage training pipeline. The first stage uses supervised fine-tuning and mixed preference optimization to internalize reasoning patterns. The second stage uses reinforcement learning with a pattern-aware reward mechanism to encourage adaptive planning and self-reflection.

Datasets

HydraFake-100K, FF++, FFIW, DF40, WILD, seeprettyface2, Talking-HeadBench, FFHQ, VFHQ, UADFV, CelebA, CelebAHQ, LFW

Model(s)

InternVL3-8B (primarily), Qwen2.5-VL-72B, Qwen3-235B-A22B, UnifiedReward-Qwen-3B, GPT-4o, Gemini-2.5-Pro, Kimi-VL-A3B-Thinking

Author countries

China

← Previous