As image synthesis improves, so does the need for reliable detectors of synthetic content. But according to a new arXiv preprint, the robustness of these detectors to transferable attacks remains incomplete and understudied. The authors argue that most existing evaluations examine only a limited set of attack conditions, leaving real-world vulnerabilities unaccounted for.
The paper proposes an approach called amortized generative attacks, which aims to craft adversarial examples that can fool a detector even when the attacker does not have full access to the target model. By amortizing the attack generation process, the method is designed to produce transferable perturbations more efficiently than traditional per-image optimization.
The study is still a preprint, and the abstract does not report specific success rates or benchmark results. Its significance lies in highlighting a gap in current detector evaluation practices and offering a new attack framework for stress-testing them. If such attacks prove effective, they could inform more robust detection methods in the future.