Learning Real Facial Concepts for Independent Deepfake Detection

View on arXiv ← Back to list

Authors: Ming-Hui Liu, Harry Cheng, Tianyi Wang, Xin Luo, Xin-Shun Xu

Published: 2025-05-07 14:31:04+00:00

AI Summary

This paper introduces RealID, a deepfake detection method that improves generalization by learning a comprehensive concept of real faces and independently assessing the probabilities of real and fake classes. RealID uses a Real Concept Capture Module and an Independent Dual-Decision Classifier to achieve state-of-the-art accuracy.

Abstract

Deepfake detection models often struggle with generalization to unseen datasets, manifesting as misclassifying real instances as fake in target domains. This is primarily due to an overreliance on forgery artifacts and a limited understanding of real faces. To address this challenge, we propose a novel approach RealID to enhance generalization by learning a comprehensive concept of real faces while assessing the probabilities of belonging to the real and fake classes independently. RealID comprises two key modules: the Real Concept Capture Module (RealC2) and the Independent Dual-Decision Classifier (IDC). With the assistance of a MultiReal Memory, RealC2 maintains various prototypes for real faces, allowing the model to capture a comprehensive concept of real class. Meanwhile, IDC redefines the classification strategy by making independent decisions based on the concept of the real class and the presence of forgery artifacts. Through the combined effect of the above modules, the influence of forgery-irrelevant patterns is alleviated, and extensive experiments on five widely used datasets demonstrate that RealID significantly outperforms existing state-of-the-art methods, achieving a 1.74% improvement in average accuracy.

Key findings

RealID significantly outperforms state-of-the-art methods across five datasets, achieving a 1.74% average accuracy improvement. The approach demonstrates robustness across different backbone architectures and effectively reduces misclassification of real images as fake. Ablation studies confirm the effectiveness of both proposed modules.

Approach

RealID uses two modules: Real Concept Capture Module (RealC2) learns real face prototypes using a Multi-Real Memory and specialized losses, and Independent Dual-Decision Classifier (IDC) makes independent decisions based on real face concepts and forgery artifacts, improving generalization.

Datasets

FF++, Celeb-DF, DFD, DFDC, DFDCp, UADFV

Model(s)

EfficientNet (primarily), also tested with Xception and ViT (ViT-L, ViT-B)

Author countries

China, Singapore

← Previous