Federated Learning

Chapter 6: Evaluating gradient inversion attacks and defenses

Yangsibo Huang^a; Samyak Gupta^b; Zhao Song^c; Sanjeev Arora^b; Kai Li^b    aElectrical and Computer Engineering, Princeton University, Princeton, NJ, United States^bPrinceton University, Princeton, NJ, United States^cAdobe Research, San Jose, CA, United States

Abstract

Gradient inversion attack (or input recovery from gradient) is an emerging threat to the security and privacy preservation of federated learning, whereby malicious eavesdroppers or participants in the protocol can partially recover the clients' private data. This chapter summarizes existing attacks in federated learning, and investigates their potential limitations. The chapter then presents an evaluation of the benefits of several ...