UC Davis

In recent years, increasing regulations aimed at protecting privacy-sensitive data have significantly restricted access to such information. To harness the vast amount of data generated across distributed entities, such as mobile devices, a decentralized approach called Federated Learning (FL) has emerged. FL allows multiple clients to collaboratively train a machine learning model while keeping all data stored locally. However, the decentralized and large-scale nature of FL introduces new challenges. Although FL shows promise in safeguarding user privacy, its security vulnerabilities pose a significant barrier to its broader adoption. Adversaries can exploit FL-specific vulnerabilities to compromise the integrity of the trained model. Additionally, the substantial communication overhead associated with FL can negate the advantages of distributing computation to the edge.In this dissertation, we first provide a comprehensive survey of the security vulnerabilities unique to FL, highlighting key attacks, defenses, and challenges. Moreover, we introduce Adaptive Federated Dropout (AFD), a framework aimed at reducing communication costs in FL, demonstrating its ability to significantly enhance efficiency and model generalization. We then explore and implement novel backdoor attacks that manipulate the global model while evading detection within the FL framework. In response to the increasing threat posed by these attacks, we test innovative defense mechanisms to safeguard FL systems. The defense strategies focus on mitigating vulnerabilities by detecting anomalous model updates and employing filtering techniques to prevent malicious contributions from impacting the global model. Finally, we explore promising future research directions to enhance the robustness of FL, mainly focusing on the next generation of defense mechanisms. Achieving end-to-end security in FL requires robust protocols capable of withstanding adversarial attacks targeting both the algorithms and the data.