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

TL;DR

The paper tackles incremental face forgery detection under evolving forgery techniques, identifying a domain-confusion challenge when replayed real samples resemble synthetic artifacts. It introduces Domain-Aware Relative Weighting (DARW), combining direct supervision on domain-safe replay samples with a Relative Separation Loss and a Domain Confusion Score to adaptively balance information preservation and artifact suppression. Using diffusion-based replay generation (LDM) with dual generators, the method achieves robust incremental detection across diverse replay settings, mitigating the adverse effects of domain overlap. Experiments across mixed-era protocols show consistent improvements over state-of-the-art baselines, with strong ablations and visualizations supporting the effectiveness of the adaptive weighting strategy in maintaining accurate decision boundaries during continual learning.

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.

Figures (9)

• Figure 1: Left: Comparison between traditional sample replay and generative replay. Right: The challenge of applying generative replay to forgery detection, as both Gen-Real and Fake are generated through similar processes, making it difficult for the detector to distinguish real from fake.
• Figure 1: Visualization of LDM-generated replay samples across multiple domains. (a–b) Real and DDPM-generated samples from the DiffusionFace dataset. (c–d) Real and Fake samples from FaceForensics++. (e–f) Real and Fake samples from DFDCP. All images are shown at a resolution of $256 \times 256$.
• Figure 2: Influence of distribution similarity on generative replay. LDM-I (ODE version song2020denoising of LDM) generates images more similar to LDM than image-level DDPM (using SDE) and DDIM. All replay generators are controlled to have comparable FID scores, ensuring similar generation quality. However, when the replay distribution is closer to the original fake distribution, detection accuracy drops, indicating that generative replay performs better when the replay generator differs more from the fake generator.
• Figure 2: Ablation on Relative Separation Loss: Sample-wise vs. Centroid-based. The Sample-wise formulation shows consistently stronger robustness on both Protocol 1 (Left) and Protocol 2 (Right), with clear advantages in the later incremental stages.
• Figure 3: The overall framework of our method.