EvolveReason: Self-Evolving Reasoning Paradigm for Explainable Deepfake Facial Image Identification

View on arXiv ← Back to list

Authors: Binjia Zhou, Dawei Luo, Shuai Chen, Feng Xu, Seow, Haoyuan Li, Jiachi Wang, Jiawen Wang, Zunlei Feng, Yijun Bei

Published: 2026-03-08 07:42:34+00:00

AI Summary

This paper introduces EvolveReason, a novel framework for explainable deepfake facial image identification that mimics human reasoning and observational processes. It constructs a Chain-of-Thought dataset, CoT-Face, to guide Vision-Language Models (VLMs) in human-like thinking and integrates a forgery latent-space distribution capture module to detect subtle high-frequency cues. EvolveReason also employs a self-evolution exploration strategy using reinforcement learning to optimize textual explanations, resulting in superior identification performance, accurate forgery detail identification, and strong generalization capabilities.

Abstract

With the rapid advancement of AIGC technology, developing identification methods to address the security challenges posed by deepfakes has become urgent. Face forgery identification techniques can be categorized into two types: traditional classification methods and explainable VLM approaches. The former provides classification results but lacks explanatory ability, while the latter, although capable of providing coarse-grained explanations, often suffers from hallucinations and insufficient detail. To overcome these limitations, we propose EvolveReason, which mimics the reasoning and observational processes of human auditors when identifying face forgeries. By constructing a chain-of-thought dataset, CoT-Face, tailored for advanced VLMs, our approach guides the model to think in a human-like way, prompting it to output reasoning processes and judgment results. This provides practitioners with reliable analysis and helps alleviate hallucination. Additionally, our framework incorporates a forgery latent-space distribution capture module, enabling EvolveReason to identify high-frequency forgery cues difficult to extract from the original images. To further enhance the reliability of textual explanations, we introduce a self-evolution exploration strategy, leveraging reinforcement learning to allow the model to iteratively explore and optimize its textual descriptions in a two-stage process. Experimental results show that EvolveReason not only outperforms the current state-of-the-art methods in identification performance but also accurately identifies forgery details and demonstrates generalization capabilities.

Key findings

EvolveReason outperforms current state-of-the-art methods in deepfake identification performance across various intra- and cross-dataset evaluations, including FF++, CelebDF, and DeepFaceGen. The framework accurately identifies forgery details and provides reliable, fine-grained textual explanations, effectively addressing issues like hallucinations common in other VLM-based approaches. Its modules (FVCE, ICA, SER) are proven effective through ablation studies, contributing to enhanced accuracy and generalization.

Approach

EvolveReason guides a Vision-Language Model (VLM) through human-like reasoning for face forgery detection by training it on a custom Chain-of-Thought (CoT-Face) dataset. It enhances visual input with a Forgery Visual Clue Extraction (FVCE) module that leverages Stable Diffusion to capture high-frequency forgery cues from the latent space. A Self-Evolving Reasoning (SER) strategy, based on reinforcement learning, iteratively refines the VLM's textual explanations for greater reliability and detail, mitigating hallucinations.

Datasets

CoT-Face, FF++ (FaceForensics++), CelebDF, DFD, DFDC, DeepFaceGen

Model(s)

Qwen-72B-VL-MAX (for CoT-Face generation and TeacherVLM), Deepseek-R1 (for refining CoT-Face), Stable Diffusion v1.5 (for FVCE), Qwen2.5-VL-7B (as base VLM)

Author countries

China

← Previous