Inference-Based Similarity Search in Randomized Montgomery Domains for Privacy-Preserving Biometric Identification

摘要

Similarity search is essential to many important applications and ofteninvolves searching at scale on high-dimensional data based on their similarityto a query. In biometric applications, recent vulnerability studies have shownthat adversarial machine learning can compromise biometric recognition systemsby exploiting the biometric similarity information. Existing methods forbiometric privacy protection are in general based on pairwise matching ofsecured biometric templates and have inherent limitations in search efficiencyand scalability. In this paper, we propose an inference-based framework forprivacy-preserving similarity search in Hamming space. Our approach builds onan obfuscated distance measure that can conceal Hamming distance in a dynamicinterval. Such a mechanism enables us to systematically design statisticallyreliable methods for retrieving most likely candidates without knowing theexact distance values. We further propose to apply Montgomery multiplicationfor generating search indexes that can withstand adversarial similarityanalysis, and show that information leakage in randomized Montgomery domainscan be made negligibly small. Our experiments on public biometric datasetsdemonstrate that the inference-based approach can achieve a search accuracyclose to the best performance possible with secure computation methods, but theassociated cost is reduced by orders of magnitude compared to cryptographicprimitives.