A Selective Privacy-Preserving Federal Learning Framework

Federated Learning (FL), as an emerging collaborative learning paradigm, has also attracted much attention for privacy preservation in FL. Existing preservation techniques usually encrypt the entire gradient, and the efficiency problem makes these schemes difficult to be applied to real life. To solve this problem, some scholars have proposed algorithms that encrypt only part of the gradient, but there are mask consensus problems as well as communication efficiency problems. To address these shortcomings, we propose a selective privacy-preserving federated learning framework, where our framework only encrypts the core partial gradient and does not require communication negotiation. Based on this, we also employ a secure aggregation algorithm to achieve defense against poisoning attacks. Security analysis shows that our approach significantly reduces the communication overhead compared to selective gradient encryption, which requires negotiation, while also effectively protecting the model from poisoning attacks.