Gradient Privacy-Preserving In Federated Learning via Proxy Re-Encryption
Yipeng Sun, Yuexiang Yang · 2022
For decentralized computing systems like cloud and edge computing, federated learning (FL) is an effective and safe machine learning technique. Its learning method involves frequent communication as participating local devices submit updates to a central server, which aggregates them and reassigns new weights to the devices. The updates may include gradients or model parameters. Private data is not sent outside of certain local devices in FL, offering a strong privacy protection solution. Since attackers can deduce the user's privacy from the local output, such as gradients, it still has some privacy problems. In order to effectively address this issue, we propose in this paper a proxy re-encryption scheme with a mask that, by incorporating a third-party proxy, effectively protects against server complicity attacks. Experiments demonstrate improved allowing the FL system to increase the acceptable communication cost while achieving more security features during model transmission. Moreover, our scheme is secure to honest-but-curious server setting even if the server colludes with multiple users. Overall, our system offers a safer and more precise system for FL.