Doctor of Philosophy (Ph.D.)
Degree Granting Department
Zhuo Lu, Ph.D.
Nasir Ghani, Ph.D.
Ismail Uysal, Ph.D.
Attila A. Yavuz, Ph.D.
Jie Xu, Ph.D.
Federated Learning, Edge Computing, Client-Selection, Poisoning Data Attack, Key Establishment
Due to the high development of wireless networking and artificial intelligence, most of the data are generated from mobile devices, which distribute in different environments. As such, how to improve the performance of machine learning-based networking and its security should be carefully considered. To reduce the communication burden and protect private information from users, Federated Learning (FL) is a possible solution for learning-based wireless networking. Although FL achieves much success until now, it also remains some specific issues to be solved. In this dissertation, we propose two FL wireless networking frameworks and discuss two potential security issues.
In the FL wireless network, due to the single server architecture, the server processes the global computing after receiving all local model updates, which means the training period high depends on the slowest clients. Because we cannot avoid all clients obtaining qualified wireless channels, it is necessary to design new FL architectures to improve the training performance from a networking perspective. Therefore, we design a new multi-server FL architecture and propose an FL algorithm for this architecture called MS-FedAvg. Next, by leveraging contextual information, we propose an online selection policy called COCS, which is based on the contextual combinatorial multi-armed bandits.
For the network security design, we first investigate the targeted model poisoning attack. To cope with this security issue, we design a new defense strategy, called LoMar, in FL based on kernel function, which can distinguish the distribution difference between honest and malicious clients. The common assumption of wireless key generation considers that the wireless channel information is sufficiently random. By leveraging the not random property, we design an efficient attack, named MLTS. Based on MLTS, we propose a design guideline for how to use the wireless channel to generate the secret key.
Scholar Commons Citation
Qu, Zhe, "Improving Wireless Networking from the Learning and Security Perspectives" (2022). USF Tampa Graduate Theses and Dissertations.