B1313
Title: Defending against backdoor attack
Authors: Yao Li - University of North Carolina at Chapel Hill (United States) [presenting]
Abstract: Backdoor attacks are getting increasing attention as studies have shown that federated learning systems can be easily fooled by them. However, defenses against such attacks are not investigated sufficiently in federated learning. Federated learning is an emerging machine learning technique as it addresses the problem of data privacy by updating models on local clients and aggregating the global model without accessing the local data. However, such distributed nature makes it vulnerable to backdoor attacks, as attackers can send malicious model updates to insert a backdoor in the global model. We study the differences between malicious updates and benign updates and propose a detection method to filter out malicious updates from attackers to protect the federated training process.
