ML Security PapersWhen Generative Replay Meets Evolving Deepfakes: Domain-Aware Relative Weighting for Incremental Face Forgery Detection
Haoyu Shen 1, Jikang Cheng 2, Renye Yan 2, Zhongyuan Wang 3, Wei Peng 4, Baojin Huang 1
Published on arXiv
2511.18436
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
DARW consistently improves incremental learning performance for face forgery detection across different generative replay settings and reduces the adverse impact of domain overlap between replay generators and new forgery models.
DARW (Domain-Aware Relative Weighting)
Novel technique introduced
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 Contributions
- Systematic characterization of generative replay for forgery detection into domain-safe and domain-risky scenarios based on generator similarity to new forgery methods
- Domain-Aware Relative Weighting (DARW) strategy with a Relative Separation Loss that balances supervision and confusion for domain-risky replay samples
- Domain Confusion Score that dynamically adjusts supervision tradeoffs based on per-sample reliability in incremental learning
🛡️ Threat Analysis
The paper's primary contribution is improving face forgery / deepfake detection — a canonical AI-generated content detection task. It proposes a novel training strategy (DARW) for keeping detectors current as new generative forgery methods emerge, qualifying as a novel detection methodology under ML09.