Data-free fingerprinting technology for biometric classifiers

Abstract

Deep neural networks (DNNs) for biometrics represent the intellectual property of model owners due to the extensive biometric data and significant computing resources required for training them. However, existing technologies for protecting the intellectual property of DNNs are primarily tailored for general tasks and often fail to account for the unique characteristics of DNNs for biometrics. Consequently, these technologies have limitations as they may increase data costs. This is because they typically require additional data samples to construct pairs that can be used for verifying intellectual property. Given the heightened privacy concerns and the challenging nature of acquiring authorized biometric data, addressing the issue of increased data costs is paramount in biometric model intellectual property protection technology. To address this challenge, we introduce MGIP (Multi-Generator Intellectual Property protection), a novel data-free fingerprinting framework specifically designed for biometric classifiers. Our key technology innovations include: (1) a collaborative multi-generator architecture that creates a variety of fingerprints without external data, (2) an adaptive threshold strategy that dynamically adjusts verification criteria, and (3) a robust fingerprint selection that ensures reliable ownership verification. In our empirical evaluation, we conduct an ablation study using three state-of-the-art technologies and six datasets, including three general datasets and three biometric datasets. Our comparative analysis demonstrates that MGIP consistently outperforms three state-of-the-art technologies in accurately identifying pirated models.