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

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 addresses incremental face forgery detection, where models must adapt to new forgery types without forgetting previous ones, proposing a novel Domain-Aware Relative Weighting (DARW) strategy. DARW systematically investigates generative replay, distinguishing between "domain-safe" and "domain-risky" generated samples. It leverages a Relative Separation Loss and a dynamically adjusted Domain Confusion Score to balance supervision and mitigate domain overlap, consistently improving incremental learning performance.

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 improves incremental learning performance for face forgery detection across various generative replay settings and effectively alleviates the negative impact of domain overlap. The method demonstrates strong robustness across different replay generators and various backbone architectures (EfficientNetB4, Xception, ResNet34). The dynamic adaptive strategy in DARW outperforms non-adaptive and fixed-weight baselines in balancing knowledge preservation and risk mitigation.

Approach

The authors propose Domain-Aware Relative Weighting (DARW) to effectively utilize generative replay for incremental face forgery detection. 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 based on sample reliability, mitigating the adverse effects of domain overlap.

Datasets

Celeb-DF-v2 (CDF), DeepFake Detection Challenge Preview (DFDCP), FaceForensics++ (FF++), SDv21, DiT, LDM, DDPM (from DiffusionFace and DF40 datasets).

Model(s)

The detection backbone used is EfficientNetB4, with Xception and ResNet34 also tested for generalization. The primary generative replay model is Latent Diffusion Model (LDM), and the Denoising Diffusion Implicit Model (DDIM) is used as the sampler during generation.

Author countries

China, USA

← Previous