Research Projects

With the increasing reliance on data-driven services, the study of information leakage has become crucial in fields where sensitive data is shared or processed. The fundamental question regarding information leakage is: ``How much information does an observation released to an adversary reveal about correlated sensitive data?'' This question arises in various applications such as healthcare, social networks, smart homes and IoT, financial services, privacy-preserving machine learning, federated learning, and location-based services. There are various operationally interpretable information leakage measures in the computer science and information theory literature, where a measure is considered operationally interpretable if an upper bound on it ensures specific real-time privacy guarantees. Each of these measures is developed to address a specific type of adversary, e.g., an adversary may be interested in maximizing the probability of correctly guessing the sensitive data, an adversary may be interested in guessing a function of the sensitive data. Depending on the adversary considered in the problem, a particular leakage measure may be more appropriate. This project aims to develop new operationally interpretable information leakage measures and apply them, alongside existing measures, to evaluate and design protocols for private information retrieval and privacy-preserving machine learning. We have three main goals of the project. (i) In our recent works, we have introduced some information leakage measures based on different types of adversaries, e.g., an adversary whose performance is measured via generalized gain functions rather than just the probability of correctly guessing, an adversary seeking to minimize the number of attempts needed to accurately guess the sensitive data, an adversary that is memoryless. We plan to expand this line of research by exploring information leakage in the presence of various types of adversaries. This will help uncover the fundamental limits of information leakage, including scenarios where the adversary is allowed to guess with a small probability of error, has limited memory, and other such constraints. (ii) In private information retrieval (PIR), a user wishes to retrieve one of the files replicated at some servers without revealing any information (i.e., perfect privacy) about the identity of the file to the servers. PIR under a relaxed notion of privacy requirement has also been studied in the literature, under the name weakly private information retrieval (WPIR). However, only few information leakage measures based on isolated types of adversaries have been considered. We plan to study WPIR by generalizing the adversary's inferential capabilities using a tunable parameter, allowing us to recover existing models as special cases, and making the results adaptable to specific applications. We also plan to investigate the fundamental limits of information leakage in generative adversarial networks.