Secure Triplet Loss for End-to-End Deep Biometrics

Abstract

Although deep learning is being widely adopted for every topic in pattern recognition, its use for secure and cancelable biometrics is currently reserved for feature extraction and biometric data preprocessing, limiting achievable performance. In this paper, we propose a novel formulation of the triplet loss methodology, designated as secure triplet loss, that enables biometric template cancelability with end-to-end convolutional neural networks, using easily changeable keys. Trained and evaluated for electrocardiogram-based biometrics, the network revealed easy to optimize using the modified triplet loss and achieved superior performance when compared with the stateof-the-art (10.63% equal error rate with data from 918 subjects of the UofTDB database). Additionally, it ensured biometric template security and effective template cancelability. Although further efforts are needed to avoid template linkability, the proposed secure triplet loss shows promise in template cancelability and non-invertibility for biometric recognition while taking advantage of the full power of convolutional neural networks.

Secure Triplet Loss: Achieving Cancelability and Non-Linkability in End-to-End Deep Biometrics

“…Considering this, we recently proposed the Secure Triplet Loss [8], a formulation of the triplet loss [30] that enables learning end-to-end models to bind biometric samples with keys. With this method, biometric templates become easily cancelable, just requiring a key change to be invalidated.…”

“…The training method proposed in [8] modifies the triplet loss to make the final sample representations cancelable (as illustrated in Fig. 1).…”

“…Considering this, in [8] we proposed the Secure Triplet Loss, a reformulation of the well-known triplet loss that enables training end-to-end deep learning models to obtain cancelable biometric templates. The proposed method allows taking full advantage of the capabilities of end-toend deep networks while still ensuring the security of the stored biometric data.…”

“…The proposed Secure Triplet Loss is reformulated to include a component that measures and actively promotes template non-linkability. The contributions of this work relative to the prior research in [8] are five-fold:…”

“…As the prior study in [8], this work includes an evaluation of identity verification performance and the template security properties of cancelability, non-linkability, noninvertibility, and secrecy leakage. For realistic results, we use the off-the-person University of Toronto ECG (UofTDB) [11] and the unconstrained YouTube Faces [12] databases, with disjoint sets of identities for training and testing.…”