Paper 2025/703

Priv-PFL: A Privacy-Preserving and Efficient Personalized Federated Learning Approach

Abstract

Federated Learning (FL) allows clients to engage in learning without revealing their raw data. However, traditional FL focuses on developing a single global model for all clients, limiting their ability to have personalized models tailored to their specific needs. Personalized FL (PFL) enables clients to obtain their customized models, either with or without a central party. Current PFL research includes mechanisms to detect poisoning attacks, in which a couple of malicious nodes try to manipulate training convergence by submitting misleading data. However, these detection approaches often overlook privacy concerns, as they require clients to share their models with all other clients. This paper extends BALANCE, a personalized poisoning detection mechanism based on client models and their expectations. Our method enhances both security and privacy by ensuring clients are not required to share their model data with other clients. By leveraging server-assisted PFL and Fully Homomorphic Encryption (FHE), we enable a central party to identify unpoisoned clients from the perspective of individual clients and train personalized models securely. Additionally, we introduce an efficient personalized client selection algorithm that prevents redundant checks and ensures the inheritance of unpoisoned clients.