Inversion, Membership, and Poisoning Attacks in Federated Learning - Challenges and Defenses

Machine-learning systems are increasingly exposed to advanced privacy and integrity threats, including model inversion, membership inference, and poisoning attacks, which can extract sensitive attributes, reveal user participation, or manipulate global models. Despite providing a decentralized training approach that keeps data on local devices, Federated Learning (FL) remains highly appealing for privacy-sensitive sectors, it remains vulnerable to these sophisticated adversarial techniques. This paper provides a focused technical analysis and classification of the three major attack categories that threaten FL, highlighting their mechanisms and practical implications for secure deployment. Our key contribution is a concise, structured examination of how these attacks operate within FL workflows and what security weaknesses they expose. The discussion is intended for researchers, practitioners, and system designers seeking to reinforce the security and dependability of federated learning frameworks.