In this tutorial, Dr. Kai Li first introduced the emerging concept of CyberEdge networks, highlighting their potential to significantly improve data integration and processing at the edge of mobile computing systems, which requires fast, private, and resilient decision-making. Then, he illustrated how federated learning could be effectively applied within CyberEdge environments, emphasizing federated learning's ability to protect sensitive data, such as patient health records and AR/VR user information, by keeping it localized on individual user devices. Dr. kai Li also provided a comprehensive survey of current adversarial attacks on federated learning and discussed corresponding defense mechanisms, paying particular attention to stealthy model feature-oriented poisoning attacks, which, despite requiring no direct data access, remain highly effective and challenging to detect.
To counter these new threats, he explored several advanced defense strategies, including Gradient-weighted Class Activation Mapping integrated with autoencoders and differential privacy-based methods. Furthermore, we outlined several promising future research directions and open challenges to enhance federated learning resilience in CyberEdge networks, particularly focusing on hybrid horizontal-vertical federated learning schemes and vulnerability analyses within resource-constrained TinyML systems.
CISTER researcher