Research Projects

Driven by Digital India and Aatmanirbhar Bharat, India has seen rapid growth in connected devices, creating a dynamic AI-powered data ecosystem with vast potential for scientific and industrial advancements. However, privacy concerns and regulatory limitations, such as India's Digital Personal Data Protection Act, restrict data sharing across multiple owners, complicating machine learning (ML) tasks. Federated learning (FL) addresses these challenges by enabling decentralized model training on edge devices, and the server is able to train a global model through secure data exchanges with the clients, while client data remains local and private. Despite significant milestones in the years, multiple fundamental challenges remain in the practical implementation of FL. Particularly, edge users have often substantial constraints on their resources (e.g., memory, compute, communication, and energy), which severely limits their capability in processing large models locally. Furthermore, the current FL systems are incompatible with the Right to be Forgotten, which empowers users to have their private information removed from ML models. Additionally, a comprehensive framework for addressing fairness concerns of the clients in FL is also missing. Hence, in this proposal, we aim to establish a foundational framework for empowering edge users with efficient, privacy-preserving, and socially fair ML solutions, ultimately contributing to the realization of a sustainable AI-driven digital future. The proposal has three major objectives. The first objective is to design fair, efficient, and optimal machine unlearning strategies, empowering users with the Right to be Forgotten. This involves creating novel strategies where a specific user can selectively and securely request data removal, without compromising the model's overall utility or fairness. The second objective, to be pursued concurrently, focuses on developing unified solutions for deploying on-device artificial intelligence at the edge. This includes enabling resource-constrained devices to collaboratively train and deploy large language models (LLMs) while maintaining privacy and efficiency. Finally, the third objective is to perform a comprehensive benchmarking analysis of the proposed methodologies on real-world biomedical datasets, particularly genomic datasets. Such datasets represent some of the most privacy-sensitive and heterogeneous data, making them ideal for evaluating the proposed solutions. The outcomes of this proposal will empower edge users with control over their data, democratize access to large AI models, and provide robust solutions for privacy-sensitive applications such as healthcare. These advancements will contribute to India's AI-driven digital transformation, aligning with the goals of Digital India and Aatmanirbhar Bharat, while addressing the critical challenges of data privacy, fairness, and resource efficiency.