Defense-in-Depth: Tamper Localization, Self-Recovery, and Attack-Type Classification for Meaningful Secret Sharing in Privacy-Preserving Face Recognition

Visual secret sharing-based face recognition divides face data into secret shares that reveal no information about the individual. The recent VOIDFace framework employs patch-based secret sharing to store training faces across various institutions and to provide users with a Right-To-Be-Forgotten (RTBF) property. Two main constraints hinder VOIDFace’s practical application. Firstly, VOIDFace shares resemble arbitrary noise and are consequently easily identifiable as ciphertext and thus prone to harvest-now-decrypt-later attacks. Secondly, VOIDFace shares lack integrity. An adversary can secretly compromise a share and corrupt the training. Even if the breach is identified, the only option would be to discard the share and retransmit it. The proposed
technique solves these problems by making the framework more secure and self-recovering. Initially, the noisy shares are embedded into cover images. This renders the ciphertext unnoticeable. Similarly, the integrity of the embedded cover images is protected with a dual layer of a fragile cryptographic hash and a semi-fragile DWT spread-spectrum watermark. The
proposed technique employs a redundant embedding scheme to reconstruct the tampered block without retransmission. Additionally, a lightweight classifier is added to read the localized tamper signature to separate malicious tampering from benign channel operations. Experiments using VGGFace2
combined with LFW, CALFW, and AgeDB-30 demonstrate indistinguishable embedding (PSNR >40dB, SSIM >0.95), robust tampering detection, self-recovery, and comparable accuracy to VOIDFace, all while preserving privacy and RTBF assurances.

You can access the full paper here.