Atul Prakash is a Professor and the Chair of the CSE Department. He works on the efficiency and robustness/security of machine learning systems and computer systems, more broadly. On the efficiency side, he and his group have been looking at data compression strategies that preserve model training effectiveness. His team also has results in the area of model sparsity. On the robustness and security side, his group has world-class expertise in the area of finding weaknesses in machine learning models. That can be used to assess weaknesses in machine learning pipelines, find safety concerns before they occur, etc.
One recent highly cited work that originated in his group shows that one can place stickers on physical traffic signs and fool state-of-the-art classifiers. That work inspired the DARPA GARD program. The traffic sign from their research was also displayed at the London Science Museum to educate visitors on blindspots of machine learning algorithms.
He and his group have also looked at security and efficiency considerations in large language models. For example, in one of their recent papers at ACL 2024 showed that the current-generation LLMs were vulnerable to being manipulated despite the use of additional LLMs to serve as safety guards.
Recent Awards:
- Distinguished Practical Paper Award, IEEE Symposium on Security and Privacy, 2016;
- Outstanding Achievement Award, EECS, 2017;
- Best Paper Award, IEEE Cybersecurity Development Conference, 2018.