51²è¹Ý

Abstract : Deep learning models have achieved super-human performance on a wide range of vision tasks; however, they remain inherently vulnerable to adversarial perturbations. In this short talk, I will present recent work that explores adversarial perturbations not just as a threat, but as a powerful tool across different deep learning domains. First, I will demonstrate their use in recovering the drift of old class prototypes in continual learning (CVPR 2024). Then, I will show how they can serve as a defense mechanism against unauthorized personalization in text-to-image diffusion models (ACM MM 2025).

About the Speaker: Dr. Sandesh Kamath is a Post Doctoral Researcher in the LAMP group led by Dr. Joost van de Weijer at Computer Vision Center(CVC) in Universitat Autonoma de Barcelona(UAB), Spain. Previously, he was a Microsoft Research Post Doctoral Fellow in Indian Institute of Technology(IIT), Hyderabad working with Prof. Vineeth N Balasubramanian. He holds a PhD from Chennai Mathematical Institute(CMI), Chennai, advised by Prof. K V Subramanyam and Dr. Amit Deshpande, Microsoft Research India and MTech from Indian Institute of Technology(IIT), Delhi, both in Computer Science. His research has been published in top-tier conferences, including NeurIPS, AAAI, and CVPR. He received the Best Paper Award (Research track) at the ACM CODS-COMAD 2022 conference and was recognized as an Outstanding Reviewer at the BMVC conference in 2024. With four years of industrial experience, he has worked in Electronic Design Automation (EDA) and Online Education. His current research focuses on Machine Learning (Deep Learning), with interests in Adversarial Robustness, Explainable AI, Continual Learning, and Generative AI.

We look forward to your active participation.