Inversion Attack Framework for Deep Face Hashing

Zihe Huang, Luyang Ying, Chuan Qin, Heng Yao, Xinpeng Zhang

Abstract: Deep face hashing enables efficient identity representation and retrieval, but the concerns of irreversible security are growing. In this letter, we propose an inversion attack framework targeting deep face hashing, aiming to reconstruct high-quality face images solely from hash codes. Specifically, A face hashing inversion network (FHINet) is presented to map hash code into spatially enhanced latent map, which is then used to guide a pre-trained StyleGAN2 generator to synthesize identity-consistent face image. A conditional discriminator is also introduced to enforce visual realism and hash alignment through adversarial training. Experimental results demonstrate that our framework reconstructs inversely face images across different deep face hashing models, achieving high visual quality and identity consistency. This work also offers a potential solution for enhancing the limited face dataset.

External IDs:dblp:journals/spl/HuangYQYZ25