Education

Thesis Title: Design and Implementation of Privacy Preserving Secure Schemes for Biometric Templates

Advisor: Prof. Sanjay Kumar Singh

Abstract: Biometrics is the study of anatomical or behavioral features of living beings for the purpose of their automatic identification. These frameworks have successfully replaced traditional token based access granting mechanisms like passwords and PIN numbers. The overwhelming success of biometric system can be attributed to the low level of security offered by the token based systems. Alternatively, biometric traits are inherently associated with some specific properties, which make them immune to such hazardous situations. A generic biometric system has numerous vulnerabilities and security issues associated with it. From intrinsic failures to adversarial attacks, biometric systems can be subjected to a wide variety of attacks. Amongst all the forms and types of these threats, attacks on template database is by far the most severe form of adversarial attack on biometric systems. A successful attack on the biometric database poses several hazardous prospects, especially for the system users. The issue of privacy is also very much relevant in this form of attack. A resourceful adversary might try to link the acquired biometric data of a user with other external sources of data, thereby mining sensitive information about the user.

This thesis is dedicated towards the development of novel schemes for countering against adversarial attacks on the biometric template database. A general solution to the problem of biometric data security is to store some appropriate data associated with the biometric samples, instead of storing the original ones. In this thesis, the notion of biometric cancelability is employed as the underlying design principle for the proposed frameworks. Under this paradigm, the original biometric signals are deliberately modified with the aid of a distortion parameter (derived from a user specific key) to produce a transformed template. The alteration process must be carried out by a specialized function which is very hard to invert under normal conditions (i.e. one-way). The essence of cancelable biometric schemes lies in the fact that the stored template can be replaced by another transformed form in case the original template gets compromised. Regeneration of the template would entirely depend on the distortion parameter.

A major problem with cancelable biometric schemes is the degradation in the overall performance of the resulting biometric recognition system. This deterioration in performance occurs since matching between two biometric templates occurs in the transformed space, rather than in their basic forms. As such, another vital goal of this dissertation is designing cancelable schemes which do not diminish the recognition accuracy rates of the biometric system. Thus these techniques would have the dual advantage of strong security guarantees along-with acceptable performance measures. The final aspect of this thesis is related to soft biometric traits. Soft biometric traits are ancillary properties which assist in the biometrics recognition process. Since there exist multiple benefits of using this information, researches in the development of fusion models incorporating soft biometric traits along-with primary ones is very active. However, the security and privacy issues which accompany them have not been addressed in the literature. This thesis concludes with a formal investigation into the privacy issues associated with soft biometric traits and subsequently attempts to design a suitable privacy preserving multimodal framework for the same.

An underlying intention of this thesis is to develop schemes based on proper cryptographic primitives and notions. The merge of biometrics and cryptography has not been successful due to the nature of the biometric signals. Biometric samples suffer from the inherent problem of intra-class variations, which make standard cryptographic techniques difficult to implement therein. However, if the merging process of cryptography and biometric is made feasible, it would introduce strong security guarantees and provable computational bounds to the system. Consequently, the various aspects of the resulting frameworks could be rigorously analyzed as well as experimentally vindicated.

Thesis Title: An Information Theoretic Framework For Differential Privacy

Advisor: Prof. Shekhar Verma

Abstract: Data or information privacy is an important aspect of today's data centric world. The principal technique which provides data privacy utilizing the data perturbation model is differential privacy. Differential privacy was extensively studied in the past, but a major contribution to it came when a comprehensive information theoretic framework was provided it. This was the first stage achieved towards the unification of privacy and information theory.

In this work we have taken the basic information theoretic framework for differential privacy and modified it to suit the same but only in a multiple query-response environment. In this model, an adversary fires multiple queries to learn about the value of a sensitive attribute corresponding to any targeted individual. In the present work, we have established a framework which provides differential privacy in this environment. Moreover we have included the concept of auxiliary side information in our framework and analyzed its effects on the end results. The main reason for doing this is because of the analogies that will be draw between the formulated framework and the famous Wyner-Ziv distributed source coding theory pertaining to information theory. This will definitely emphasize the connection between information theory and data privacy.

To explain in brief, two frameworks are developed in this work. The first framework is used to analyze the effects of addition of Laplacian noise (the main ingredient for providing differential privacy) to the original data. The second one is the main work where differential privacy is provided in a multiple query response environment. The frameworks are successfully simulated in MATLAB and the characteristics of certain parameters relating to the frameworks are analyzed.