Facial image watermarking is often used to trace and certify images in biometric pipelines, but the authors of a new paper argue that provenance or manipulation detection alone does not guarantee reliable biometric identity. They introduce a watermark-preserving face evasion attack that targets this gap, operating under a black-box watermarking setting without access to the watermarking components.

The proposed PGD-L2 attack guides adversarial perturbations with spatial and spectral preservation terms. This reduces face-recognition similarity while limiting pixel-level distortion and high-frequency changes that could interfere with hidden watermark cues. The attack was evaluated on controlled facial images using two watermarking schemes, two face-recognition models, and two datasets.

Results show evasion rates above 92% at operational thresholds, while watermark recovery remains above 93% for StegaFormer and above 99.7% for FaceSigns. Visual quality stays above 47dB PSNR at epsilon_2=2.0. The authors say these findings expose a structural gap in watermark-based facial certification: a face image can remain watermark-valid and visually faithful while becoming biometrically unreliable.