SecureLiteFL: A Secure and Bandwidth-Saving Federated Learning Framework for Aggregation Resistant to Collusion Attacks

Federated Learning (FL) enables collaborative model training by sharing gradients while preserving data privacy. However, existing schemes lack effective verification mechanisms against server-user collusion. We propose SecureLiteFL, a verifiable and communication-efficient FL framework resistant to collusion attacks. SecureLiteFL integrates an optimized doublemasking algorithm and an ECC-based commitment scheme to protect gradient privacy and enable efficient verification. Experiments show SecureLiteFL achieves $57 \%$ faster training and $18.2 \%$ lower client-side communication overhead than VerifyNet, with verification costs independent of gradient dimensions.