異常検知における汎化性および適応性の課題について【AIセキュリティ＆プライバシーチーム】

Abstract

Anomaly detection is a critical task for ensuring the reliability, safety, and security of machine learning applications. Despite significant progress in this area, several key challenges remain unresolved—particularly in terms of generalization, adaptability to new datasets, and universality across domains. In this talk, I will explore these ongoing challenges and present our proposed solutions aimed at addressing them. Our approach focuses on enhancing robustness and reliability in anomaly detection systems, with the goal of improving their practical deployment in diverse real-world scenarios.

Bio

Mohammad Sabokrou is a Staff Research Scientist at the Machine Learning and Data Science (MLDS) Unit of the Okinawa Institute of Science and Technology (OIST). His research sits at the intersection of computer vision and trustworthy AI, with a focus on anomaly and out-of-distribution detection, continual learning, and machine learning robustness. Before OIST, he held academic positions at the Institute for Research in Fundamental Sciences (IPM) in Tehran and conducted postdoctoral research at institutions across Finland and France. He regularly contributes to top-tier ML conferences (e.g., CVPR, ICLR, NeurIPS) and serves as an area chair at ICLR 2025.