Privacy-preserving techniques in FL.

Abstract

Federated Learning (FL) has the potential to train models on decentralized data while maintaining data privacy. However, trust and security are major concerns in federated learning architecture due to the possibility of malicious contributors. This essay aims to improve the longevity of the federated learning system by addressing trust-based mitigation. The study explores different types of trust, including contributor trustworthiness, model trustworthiness and data bias. In addition, it examines trust-based mitigation techniques for the FL system, including reputation-based modelling, secure aggregation protocols and privacy-preserving techniques. These mechanisms recognise and deal with the non-trusted participants, ensuring the integrity of the FL system. The study also investigates the effects of trust-based mitigation techniques on the performance and efficiency of FL systems, balancing security measures and computational load. The approach is tested through experiments and simulations using real-world datasets and scenarios, estimating its performance in terms of model accuracy, convergence rate, communication efficiency and resemblance to application scenarios against adversarial attacks. This essay contributes to the field of FL by addressing trust challenges and providing effective mitigation strategies, paving the way for a more secure and reliable FL system in sensitive and private domains such as healthcare, finance and smart cities.