When Generative Replay Meets Evolving Deepfakes: Domain-Aware Relative Weighting for Incremental Face Forgery Detection

View on arXiv ← Back to list

Authors: Hao Shen, Jikang Cheng, Renye Yan, Zhongyuan Wang, Wei Peng, Baojin Huang

Published: 2025-11-23 13:09:02+00:00

AI Summary

This paper introduces the Domain-Aware Relative Weighting (DARW) strategy to enable effective generative replay for incremental face forgery detection (IFFD). DARW classifies generated samples as safe or risky, using a dynamic Domain Confusion Score (DC Score) to modulate a Relative Separation Loss (RS Loss) for optimal knowledge preservation. Experiments show DARW significantly improves incremental performance and robustness against domain overlap compared to existing replay methods.

Abstract

The rapid advancement of face generation techniques has led to a growing variety of forgery methods. Incremental forgery detection aims to gradually update existing models with new forgery data, yet current sample replay-based methods are limited by low diversity and privacy concerns. Generative replay offers a potential solution by synthesizing past data, but its feasibility for forgery detection remains unclear. In this work, we systematically investigate generative replay and identify two scenarios: when the replay generator closely resembles the new forgery model, generated real samples blur the domain boundary, creating domain-risky samples; when the replay generator differs significantly, generated samples can be safely supervised, forming domain-safe samples. To exploit generative replay effectively, we propose a novel Domain-Aware Relative Weighting (DARW) strategy. DARW directly supervises domain-safe samples while applying a Relative Separation Loss to balance supervision and potential confusion for domain-risky samples. A Domain Confusion Score dynamically adjusts this tradeoff according to sample reliability. Extensive experiments demonstrate that DARW consistently improves incremental learning performance for forgery detection under different generative replay settings and alleviates the adverse impact of domain overlap.

Key findings

DARW consistently achieves superior incremental detection accuracy, scoring the highest average AUC across challenging protocols (e.g., 0.9574 AUC on Protocol 1 T6), outperforming leading sample replay and continual learning baselines. The dynamic weighting strategy significantly reduces catastrophic forgetting, demonstrating a lower performance dropping rate (PD) than fixed-alpha or non-adaptive generative replay methods. Ablation studies confirm that both the Gen-Real supervision and the Relative Separation Loss are crucial for mitigating domain confusion in the latent space.

Approach

The method utilizes generative replay via the LDM architecture to synthesize past data for continual learning. It identifies domain-risky samples (where generated real samples blur the domain boundary) and applies a Relative Separation Loss (RS Loss) to enforce feature separation between generated real and fake samples. This separation is dynamically weighted by a Domain Confusion Score (DC Score) which measures domain distance, ensuring adaptive knowledge retention while mitigating confusion.

Datasets

Celeb-DF-v2 (CDF), DeepFake Detection Challenge Preview (DFDCP), FaceForensics++ (FF++), SDv21, DiT, LDM (from DiffusionFace), DDPM (from DiffusionFace)

Model(s)

EfficientNetB4 (detector backbone), LDM (Latent Diffusion Model) (generator)

Author countries

China, USA

← Previous