RAIDX: A Retrieval-Augmented Generation and GRPO Reinforcement Learning Framework for Explainable Deepfake Detection

Authors: Tianxiao Li, Zhenglin Huang, Haiquan Wen, Yiwei He, Shuchang Lyu, Baoyuan Wu, Guangliang Cheng

Published: 2025-08-06 15:08:16+00:00

AI Summary

RAIDX introduces a novel deepfake detection framework that synergizes Retrieval-Augmented Generation (RAG) and Group Relative Policy Optimization (GRPO) to enhance both detection accuracy and decision explainability. It leverages RAG for integrating external knowledge and GRPO for autonomously generating fine-grained textual explanations and saliency maps without extensive manual annotations. Experiments show RAIDX achieves state-of-the-art detection performance and provides interpretable rationales in both textual descriptions and saliency maps.

Abstract

The rapid advancement of AI-generation models has enabled the creation of hyperrealistic imagery, posing ethical risks through widespread misinformation. Current deepfake detection methods, categorized as face specific detectors or general AI-generated detectors, lack transparency by framing detection as a classification task without explaining decisions. While several LLM-based approaches offer explainability, they suffer from coarse-grained analyses and dependency on labor-intensive annotations. This paper introduces RAIDX (Retrieval-Augmented Image Deepfake Detection and Explainability), a novel deepfake detection framework integrating Retrieval-Augmented Generation (RAG) and Group Relative Policy Optimization (GRPO) to enhance detection accuracy and decision explainability. Specifically, RAIDX leverages RAG to incorporate external knowledge for improved detection accuracy and employs GRPO to autonomously generate fine-grained textual explanations and saliency maps, eliminating the need for extensive manual annotations. Experiments on multiple benchmarks demonstrate RAIDX's effectiveness in identifying real or fake, and providing interpretable rationales in both textual descriptions and saliency maps, achieving state-of-the-art detection performance while advancing transparency in deepfake identification. RAIDX represents the first unified framework to synergize RAG and GRPO, addressing critical gaps in accuracy and explainability. Our code and models will be publicly available.

Key findings

RAIDX achieved state-of-the-art detection accuracy on the SID-Set benchmark and demonstrated strong zero-shot generalization to 18 unseen generative models. The GRPO module significantly improved the quality of fine-grained textual explanations and saliency maps, providing superior interpretability. Additionally, RAIDX showed robust performance against various visual perturbations like JPEG compression and Gaussian blur.

Approach

RAIDX integrates a Vision Transformer (ViT) for feature extraction, a Retrieval-Augmented Generation (RAG) module to incorporate external knowledge from similar images, and a partially trainable Large Language Model (LLM) with LoRA adapters. The LLM is optimized using Group Relative Policy Optimization (GRPO) to generate fine-grained textual explanations and saliency maps, providing interpretable deepfake detection.

Datasets

SID-Set, COCO, Flickr, Stable Diffusion (v1.4, v1.5, v2.0, v2.1), DALLE 2, Midjourney v5, IF, GLIDE, SDXL, DALLE-3, DiffusionDB, GLIDE Stylization, playground v2.5, DFDC, LaMa, SD2IP, SD2SR

Model(s)

UNKNOWN

Author countries

United Kingdom, China, Hong Kong

← Previous