Athena Seminar Series: Cryptography, Security and Machine Learning

Robust Machine Learning (ML) is arguably the most important technical challenge of current times, to address growing concerns about misuse of AI, violations of data privacy and stealing of trained models. The problem is only exacerbated by lack of explainability for Machine Learning decisions juxtaposed with tremendous rate of adoption of AI across all industries. In this talk, I will narrate a few research threads pursued in our group. First, I will present a theoretical understanding of adversarial attacks and countermeasures inspired by the same. Second, a case study with the intelligent perception module of a (semi-) autonomous vehicle will be discussed. In that, we will see how various techniques for enhancing robustness of an ML accelerator can be stepwise integrated. The last part of the talk will highlight two new directions in ML, first, deeper understanding of the limitations of LLM and second, how robustness of ML can be achieved by borrowing techniques from cryptography.