COSIC Seminar "When AI Leaves the Lab: Security, Robustness, and..." (Alireza Aghabagherloo, COSIC) скачать в хорошем качестве

COSIC Seminar "When AI Leaves the Lab: Security, Robustness, and..." (Alireza Aghabagherloo, COSIC)

COSIC Seminar - When AI Leaves the Lab: Security, Robustness, and Trust in Modern AI Systems - Alireza Aghabagherloo (COSIC) As AI systems transition from laboratory benchmarks to real-world deployment, ensuring their security and robustness becomes a critical concern. Despite achieving high accuracy, modern deep learning models remain vulnerable to adversarial attacks, data poisoning, and privacy threats. This presentation examines the origins of these vulnerabilities by analyzing robust, non-robust, and illusory robust features in neural networks. We demonstrate why traditional robustification methods can provide a false sense of security and highlight how feature purification can meaningfully improve robustness. The talk also explores the role of data quality and duplication in shaping learning dynamics. It extends these insights to advanced settings, including reinforcement learning, federated learning, and large language and vision-language models. Attendees will gain a nuanced understanding of AI vulnerabilities and practical strategies to enhance model reliability in real-world environments.