Face-D(^2)CL: Multi-Domain Synergistic Representation with Dual Continual Learning for Facial DeepFake Detection

Authors: Yushuo Zhang, Yu Cheng, Yongkang Hu, Jiuan Zhou, Jiawei Chen, Yuan Xie, Zhaoxia Yin

Published: 2026-04-09 12:18:42+00:00

AI Summary

This paper introduces Face-D(^2)CL, a novel framework for facial DeepFake detection designed to address insufficient feature representation and catastrophic forgetting in continual learning scenarios. It achieves this by leveraging multi-domain synergistic representation to fuse spatial and frequency-domain features, and a dual continual learning mechanism combining Elastic Weight Consolidation (EWC) and Orthogonal Gradient Constraint (OGC). This replay-free approach ensures robust anti-forgetting capabilities and agile adaptability to evolving facial forgery patterns.

Abstract

The rapid advancement of facial forgery techniques poses severe threats to public trust and information security, making facial DeepFake detection a critical research priority. Continual learning provides an effective approach to adapt facial DeepFake detection models to evolving forgery patterns. However, existing methods face two key bottlenecks in real-world continual learning scenarios: insufficient feature representation and catastrophic forgetting. To address these issues, we propose Face-D(^2)CL, a framework for facial DeepFake detection. It leverages multi-domain synergistic representation to fuse spatial and frequency-domain features for the comprehensive capture of diverse forgery traces, and employs a dual continual learning mechanism that combines Elastic Weight Consolidation (EWC), which distinguishes parameter importance for real versus fake samples, and Orthogonal Gradient Constraint (OGC), which ensures updates to task-specific adapters do not interfere with previously learned knowledge. This synergy enables the model to achieve a dynamic balance between robust anti-forgetting capabilities and agile adaptability to emerging facial forgery paradigms, all without relying on historical data replay. Extensive experiments demonstrate that our method surpasses current SOTA approaches in both stability and plasticity, achieving 60.7% relative reduction in average detection error rate, respectively. On unseen forgery domains, it further improves the average detection AUC by 7.9% compared to the current SOTA method.

Key findings

Face-D(^2)CL significantly outperforms current SOTA approaches in both stability and plasticity, achieving a 60.7% relative reduction in average detection error rate. It also improves the average detection AUC by 7.9% on unseen forgery domains. The framework demonstrates strong robustness to task order variations and various image degradations like dropout, shuffle, Gaussian noise, and median blur.

Approach

The method solves the problem by integrating a multi-domain synergistic representation that fuses spatial, wavelet, and Fourier domain features to capture diverse forgery traces comprehensively. This is combined with a dual continual learning mechanism consisting of class-aware Elastic Weight Consolidation (EWC) for global parameter stability and Orthogonal Gradient Constraint (OGC) for flexible low-rank adaptation, enabling adaptation without historical data replay.

Datasets

Celeb-DF v2 (CDF2), Deepfake Detection Challenge Preview (DFDCP), Deepfake Detection (DFD), FaceForensics++ (FF++), MCNet, BlendFace, StyleGAN3 (from DF40), DF40, UADFV, WildDeepfake

Model(s)

CLIP ViT-L/14 (backbone), LoRA (Low-Rank Adaptation) modules, two-layer classifier

Author countries

China

← Previous