Research Projects

AI has significant economic implications for India, particularly for achieving Atma-nirbhar Bharat. As AI is used in mission-critical domains like defense and health, security becomes vital. Security attacks can be harmful to surveillance cameras, toll posts, and self-driving cars. DeepFake media can be misused to malign VIPs and inflict social harm. Previous AI research in India has mainly focused on accuracy, but security has been largely ignored. The objective is to test the vulnerability of AI models against various attacks, such as patch-based and adversarial input attacks, fault-injection attacks, and targeted misclassification attacks. The project will test security of well-known models such as vision-transformer, graph neural network, and natural-language processing models. Techniques will be proposed to make it difficult for adversaries to degrade DNN accuracy, providing multiple avenues for protection. The project will also evaluate ensembling techniques, where multiple DNNs are used in parallel, to provide stronger security. Retraining-free techniques will be proposed for patch-based attacks. The project will fundamentally change the DNN design paradigm by proposing methodologies that create secure-by-design DNNs. These DNN architectures will be robust to class-imbalance, out-of-distribution data, fake data, and perturbations. The secure AI algorithms and accelerators designed by this project will enhance citizens' confidence in AI services and foil attacks by malicious inland/foreign entities.

0