RetFL: a privacy-preserving and traceable framework for robust federated learning

Federated learning (FL) enables multiple clients to jointly train a model without sharing raw data. Decentralized federated learning (DFL) further removes the need for a trusted central coordinator in the aggregation process. However, in decentralized settings, model aggregation is vulnerable to inference and poisoning attacks, and achieving efficient and traceable training remains challenging. To address these challenges, we propose RetFL, a CKKS-enabled robust aggregation framework for DFL. Specifically, we establish a decentralized training workflow with VRF-based candidate selection and view change. Then, we design a weighted aggregation scheme that incorporates cosine similarity and a dynamic reputation mechanism to weight updates and suppress persistently malicious participants. Finally, we enable scalable encrypted aggregation by efficiently realizing normalization verification within CKKS. Experimental results indicate that RetFL maintains performance close to standard FL methods even under challenging adversarial conditions, while achieving better model quality than existing robust FL approaches under comparable robustness requirements.