DDCM: Cracking Anonymized Facial Images Using Denoising Diffusion Cryptanalytic Model

Currently, in consumer electronics industry, quite a number of E-healthcare products highly depend on consumers’ facial biometrics for both disease diagnosis and identity authentication. In light of this, various facial privacy-preserving approaches oriented towards the Internet of medical things have been proposed, including visual cryptography and desensitization technique. Nevertheless, few attentions have been paid to investigating their countermeasure, i.e., facial privacy analysis, owning to significant distribution discrepancy between the original and anonymized facial images. To this end, a general Denoising Diffusion Cryptanalytic Model (DDCM) is proposed in this paper to reconstruct the original facial images in a coarse-to-fine manner from their encrypted ones with diverse anonymous approaches. Specifically, to accurately reconstruct the original facial image without any prior knowledge, a customized initial prediction module with an encoder-decoder structure is introduced by taking advantage of its leaked visual cues to generate the coarse face. Furthermore, the residual map between the original and coarse facial images is explicitly modeled with the output of the customized prediction module as a conditional prior, thereby obtaining a refined facial image. Finally, extensive experiments demonstrate that the proposed DDCM could not only effectively crack multiple handcrafted facial anonymization approaches, such as blurring, pixelation and thumbnail-preserving encryption, but also outperform the previous works (e.g., Pix2Pix) in terms of image fidelity and attack success rate.