SoK: Gradient Inversion Attacks in Federated Learning

Explore a comprehensive systematization of knowledge (SoK) conference presentation examining gradient inversion attacks in federated learning environments. Learn how these critical security threats exploit client model updates to reconstruct private training data, thereby compromising participant privacy in collaborative machine learning systems. Discover the various threat models that define adversary knowledge and capabilities for executing these attacks, and understand a systematic taxonomy that categorizes gradient inversion attacks while providing practical insights into their methods and real-world applicability. Examine defensive mechanisms specifically designed to mitigate these privacy-breaking attacks, and explore the evaluation metrics used to measure attack success and assess model vulnerability. Gain insights into the key challenges facing the field and promising future research directions based on thorough analysis of existing literature, presented by researchers from the University of Salerno at the USENIX Security '25 conference.