Multibiometric Secure System Based on Deep Learning (1708.02314v1)

Abstract: In this paper, we propose a secure multibiometric system that uses deep neural networks and error-correction coding. We present a feature-level fusion framework to generate a secure multibiometric template from each user's multiple biometrics. Two fusion architectures, fully connected architecture and bilinear architecture, are implemented to develop a robust multibiometric shared representation. The shared representation is used to generate a cancelable biometric template that involves the selection of a different set of reliable and discriminative features for each user. This cancelable template is a binary vector and is passed through an appropriate error-correcting decoder to find a closest codeword and this codeword is hashed to generate the final secure template. The efficacy of the proposed approach is shown using a multimodal database where we achieve state-of-the-art matching performance, along with cancelability and security.

• The paper introduces a multibiometric framework that leverages deep neural networks and error-correction coding to generate cancelable biometric templates.
• It fuses face and iris features through fully connected and bilinear architectures, achieving a 99.7% genuine accept rate at a 100-bit security level.
• The system minimizes privacy leakage and resists unauthorized access without relying on additional side information, ensuring robust security.

Multibiometric Secure System Based on Deep Learning

The paper introduces a sophisticated multibiometric secure system that leverages deep learning and error-correction coding to provide robust and secure biometric authentication. This research aims to address the inherent privacy concerns of multibiometric systems, which store multiple biometric traits for each user, by generating cancelable and secure multibiometric templates.

The proposed framework employs deep neural networks (DNNs) to extract features from the face and iris biometrics and subsequently fuses these features into a shared multibiometric representation. Two architectures are explored for feature fusion: a fully connected architecture (FCA) and a bilinear architecture (BLA). These architectures enable the development of a robust multibiometric shared representation, which serves as the basis for generating a cancelable biometric template using feature selection processes. This template is encoded into a binary vector, which is then passed through an appropriate error-correction decoder—specifically a Reed-Solomon (RS) decoder—allowing for the generation of secure multibiometric sketches.

The method demonstrates efficacy on a multimodal database, achieving high genuine accept rates (GAR) while maintaining a secure level defined by the number of bits, known as security level. For example, with a symbol size of 6 in Reed-Solomon coding, the authors report a GAR of 99.7% at a security level of 100 bits, which is considered highly challenging to breach using brute force attacks. The system's resilience against unauthorized access (reflected in the false accept rate, FAR) is directly related to the security parameter $k$, where $FAR \approx 2^{-k}$.

The novel integration of error-coding frameworks like secure sketch and fuzzy commitment within the biometric domain marks a significant contribution, claiming to address security without necessitating additional side information like a syndrome or saved message keys, which are usually critical in similar security arrangements.

Security analysis reveals that the design minimizes privacy leakage as reflected by the mutual information estimate. Even with partial exposure to the secure sketch or user-specific data keys, the computational burden to deduce the original biometric feature vector remains prohibitively high, thereby fortifying against potential adversaries.

This research offers several practical implications. As outlined, the design can considerably enhance the operational security of applications requiring stringent access control mechanisms. The cancelable and secure templates ensure that user privacy is proactively protected even if biometric data is intercepted. Theoretically, it opens avenues for sophisticated DNN implementations that facilitate cross-modal biometric integration, complementing other emerging secure biometric solutions.

Future studies could further refine the deep learning models, particularly in fine-tuning various CNN layers to boost efficiency and accuracy. Another prospective development could involve adapting this framework for real-time application scenarios, where rapid authentication is paramount. Additionally, other biometric modalities could be introduced, thereby testing the scalability and flexibility of deep learning-based fusion techniques across more varied datasets.

Overall, the framework provides a robust foundation upon which future multibiometric systems can be designed, paving the way for more secure, reliable, and efficient biometric verification systems.