Enhancing Biometric-Capsule-based Authentication and Facial Recognition via Deep Learning

Abstract

In recent years, developers have used the proliferation of biometric sensors in smart devices, along with recent advances in deep learning, to implement an array of biometrics-based authentication systems. Though these systems demonstrate remarkable performance and have seen wide acceptance, they present unique and pressing security and privacy concerns. One proposed method which addresses these concerns is the elegant, fusion-based BioCapsule method. The BioCapsule method is provably secure, privacypreserving, cancellable and flexible in its secure feature fusion design. In this work, we extend BioCapsule to face-based recognition. Moreover, we incorporate state-of-art deep learning techniques into a BioCapsule-based facial authentication system to further enhance secure recognition accuracy. We compare the performance of an underlying recognition system to the performance of the BioCapsuleembedded system in order to demonstrate the minimal effects of the BioCapsule scheme on underlying system performance. We also demonstrate that the BioCapsule scheme outperforms or performs as well as many other proposed secure biometric techniques. CCS CONCEPTS • Security and privacy → Biometrics; Privacy-preserving protocols.

Towards Protecting Face Embeddings in Mobile Face Verification Scenarios

“…These methods use some sort of algorithmically defined transformation to convert the face embeddings (learned from the input face images) to more secure representations. The proposed transformations include one-way cryptographic [6] or Winner Takes All [7] hashing, convolution of the embedding with a random kernel [8], use of the Fuzzy Commitment [9] or Fuzzy Vault [10] scheme, fusion of a subject's face embedding with a different subject's face embedding using keys extracted from the two sets of features [11], and homomorphic encryption [12]. The main issue with these approaches is that they have not been comprehensively evaluated in terms of their ability to simultaneously satisfy all three properties of face embedding protection methods.…”

“…It is, therefore, precisely these embeddings that the attacker in our irreversibility analysis is trying to recover. This means that the attacker should not have access to these embeddings; however, they may be assumed to have access to the development set, since these embeddings are not used for enrollment 11 . The attacker could, therefore, use the development set's reference face embeddings to estimate the distribution of the 128 elements in the evaluation set's reference embeddings, which they are attempting to recover.…”

Biometric Template Protection for Neural-Network-based Face Recognition Systems: A Survey of Methods and Evaluation Techniques

“…Among the Non-NN face BTP methods considered thus far, hashing appears to be the most popular approach; however, alternatives are proposed in [19], [30], [31]. In [19], face features extracted from different face regions (using a CNN) are fused, and the fused feature vector is convolved with a random kernel.…”

“…In [19], face features extracted from different face regions (using a CNN) are fused, and the fused feature vector is convolved with a random kernel. A fusion approach is also adopted in [30], but this time a DNN-extracted face feature vector is fused with a different person's face feature vector. Specifically, a key is generated from each feature vector, then each key is used to alter the other feature vector via element-wise multiplication, followed by vector addition of the altered feature vectors.…”

“…This approach was used in [6], [13], [14], [18], [19]. Similarly, [11] trained a DNN to reduce a face feature vector to its "intrinsic" dimensionality before encrypting it (via HE), while [21] and [30] trained a CNN and an SVM, respectively, to classify already-protected face templates. The second common reason for using a train dataset among Non-NN face BTP methods, was to optimise certain parameters of handcrafted BTP algorithms (which could then be applied to the test dataset) or to train machine-learning BTP methods, but not to train a neural network.…”