ML-FAS: Multi-Level Face Anonymization Scheme and Its Application to E-Commerce Systems

Donghua Jiang, Jawad Ahmad, Zhufeng Suo, Mashael M. Alsulami, Yazeed Yasin Ghadi, Wadii Boulila

Abstract: With the proliferation of electronic commerce, the facial data used for identity authentication and mobile payment are potentially subject to data analytics and mining attacks by third-party platforms, which has raised public privacy concerns. To tackle the issue, a novel Multi-Level Face Anonymization Scheme (ML-FAS) based on deep learning technology is proposed in this paper. Firstly, a 4-D chaotic system is employed to construct different levels of keys with initial parameters securely distributed and managed using the Semiconductor SuperLattice Physical Unclonable Function (SSL-PUF). Secondly, under the guidance of the known prior distribution and adversarial training strategy, a noise-like cipher image is generated by the encryption network to withstand the known-plaintext attacks. Besides, different levels of recipients can leverage the identical decryption network to reconstruct the facial images with varying visual content. Compared with the existing manually designed anonymization schemes, the ML-FAS possesses several significant merits. Finally, extensive simulation experiments verified the effectiveness of the proposed scheme, including its security and robustness. The code is available at https://github.com/DonghuaJiang/MLFAS.