Dr. Shucheng Yu is an Associate Professor of Electrical and Computer Engineering at Stevens Institute of Technology, where he directs the Analytics and Information Security for Complex Systems Lab (AISecLab). He received his PhD in Electrical and Computer Engineering from Worcester Polytechnic Institute in 2010. His research interest is on cybersecurity in general, with recent focuses on information security, applied cryptography, wireless networking and sensing, distributed trust, applied machine learning, and practical security and privacy in IoT systems. He is the recipient of the Test of Time Paper Award of IEEE Infocom 2020. He is a Fellow of IEEE and AAIA.
Title: Federated Learning Privacy and Robustness: Cryptographic Perspectives and Wireless PerspectivesÂ
Abstract: Federated Learning (FL) has been a trendy privacy-enhancing technique in numerous distributed systems. In FL, data remains at local devices during the training process. However, data privacy can still be breached under malicious attacks including data reconstruction or membership inference attacks. For better privacy protection, privacy-protection FL (PPFL) is desired in which local parameters are encrypted or obfuscated before uploading. However, PPFL makes model poisoning attacks more convenient. On the one hand, PPFL requires local model parameters to be confidential; on the other hand, FL robustness is usually based on anomaly detection mechanisms which are usually conducted on plaintext models. Therefore, achieving PPFL and FL robustness simultaneously is a challenge. This talk delves into this problem with both cryptographic approaches and wireless approaches.