GenPure: Foundation-Model-Guided Multi-Stage Purification Framework for Black-Box Evasion of Deepfake Detectors

Although deepfake detection methodologies have advanced considerably, they remain susceptible to evasion techniques, a critical vulnerability in forensic systems. Most detection approaches that rely on generator-specific artifacts or high-frequency fingerprints become susceptible to circumvention when these discriminative features are eliminated. As a means to expose such blind spots in state-of-the-art forensics, we introduce GenPure, a novel multi-stage purification framework that systematically removes both low- and high-level synthetic cues, enabling generated images to bypass diverse detectors under a strict black-box setting. In the GenPure detection evasion pipeline, a U-Net-based reconstruction module with multi-scale, spatially adaptive Gaussian kernels first attenuates synthetic artifacts while preserving perceptual fidelity. Next, a vision foundation model—intentionally unbiased toward authentic-synthetic classification—aligns the attenuated images with authentic image characteristics by minimizing distribution discrepancies in latent space. Finally, reference-guided statistical color recalibration restores natural tone and further masks residual forensic cues. The framework requires no detector knowledge or additional training and generalizes across generator types, resolutions, and content domains. Extensive experiments demonstrated state-of-the-art evasion performance across multiple unseen generative sources and detection systems without degrading visual quality. The results exposed persistent structural weaknesses in current forensic pipelines and underscore the need for semantically grounded, manipulation-invariant defenses.