Sanitizing Diffusion-Generated Images via Fingerprint Removal and Adversarial Perturbation for Forensic Evasion

Diffusion models have rapidly advanced the realism of synthetic image generation, posing new challenges for forensic detectors. This paper proposes a two-stage forensic evasion framework designed to undermine the detectability of diffusion-generated images. In the first stage, a spectrum-aware generative adversarial network (GAN) removes frequency-domain fingerprints that are commonly exploited by forensic models. In the second stage, adversarial perturbations are applied using the Iterative Fast Gradient Sign Method (I-FGSM) to further mislead detectors while preserving visual fidelity. Experiments conducted on COCO-based datasets demonstrate that our method significantly reduces detection accuracy across multiple state-of-the-art forensic models, including UniFD, DIGBD, and SSIP. Furthermore, we show that combining fingerprint removal with adversarial perturbation achieves stronger evasion than either method alone. Ablation studies also highlight the benefits of adaptive perturbation strengths and data augmentation for enhancing cross-model evasion. This work reveals critical vulnerabilities in current forensic approaches and underscores the need for more robust detection systems against adaptive evasion strategies.