Security Threats and Defence Mechanisms in Federated Learning: A Comprehensive Review

Federated Learning (FL) is a promising decentralised machine learning model that enables multiple devices to collaboratively train a shared model without sharing their private data. While this approach enhances data privacy and regulatory compliance, it is significantly vulnerable to a range of security threats and adversarial attacks. This research seeks to investigates various attack vectors in FL, such as poisoning attacks, Byzantine attacks, Sybil attacks, and gradient inversion and also evaluates their impact on model performance and data confidentiality. Through a comprehensive analysis and empirical reviews of existing literature, the study explores mitigation strategies, attack model and threat taxonomy to classify adversarial behaviours. Key findings from the reviews suggests that while existing defence mechanisms show promise, they often suffer from trade-offs between model accuracy, system scalability, and computational overhead. The study was concluded by identifying gaps in current literature, such as the need for adaptive mitigation strategies and more realistic threat models, and offers recommendations for future work. By addressing these challenges, the research strengthens the robustness and trustworthiness of federated learning systems in real-world applications.