Making DeepFakes More Spurious: Evading Deep Face Forgery Detection via Trace Removal Attack

DeepFakes are raising significant social concerns. Although various DeepFake detectors have been developed as forensic countermeasures, these still vulnerable to attacks. Recently, a few attacks, principally adversarial succeeded in cloaking images evade detection. However, attacks typical detector-specific designs, which require prior knowledge about the detector, leading poor transferability. Moreover, only consider simple security scenarios. Less is known how effective they high-level scenarios where detector's defensive capability or attacker's varies. To address challenges, this paper, we propose novel attack for anti-forensics called trace removal attack. Instead of investigating detector side, looks into original creation pipeline, attempting remove all detectable natural traces render fake more “authentic”. This detector-agnostic design allows be against arbitrary even unknown detectors. implement attack, first perform an in-depth discovery, identifies three discernible traces: spatial anomalies, spectral disparities, and noise fingerprints. Then learning-based network (TR-Net) proposed that involves one generator multiple discriminators. Each discriminator responsible individual representation avoid cross-trace interference. These discriminators arranged parallel, prompts simultaneously. evaluate efficacy crafted heterogeneous were embedded with different levels defense, attackers' background data varied. The experimental results show can significantly compromise detection accuracy six state-of-the-art while causing negligible degradation visual quality.