International Association
for Cryptologic Research

Federated Learning (FL) enables collaborative model training while preserving data privacy by avoiding the sharing of raw data. However, in large-scale FL systems, efficient secure aggregation and dropout handling remain critical challenges. Existing state-of-the-art methods, such as those proposed by Liu et al. (UAI'22) and Li et al. (ASIACRYPT'23), suffer from prohibitive communication overhead, implementation complexity, and vulnerability to poisoning attacks. Alternative approaches that utilize partially connected graph structures (resembling client grouping) to reduce communication costs, such as Bell et al. (CCS'20) and ACORN (USENIX Sec'23), face the risk of adversarial manipulation during the graph construction process.

To address these issues, we propose ClusterGuard, a secure clustered aggregation scheme for federated learning. ClusterGuard leverages Verifiable Random Functions (VRF) to ensure fair and transparent cluster selection and employs a lightweight key-homomorphic masking mechanism, combined with efficient dropout handling, to achieve secure clustered aggregation. Furthermore, ClusterGuard incorporates a dual filtering mechanism based on cosine similarity and norm to effectively detect and mitigate poisoning attacks.

Extensive experiments on standard datasets demonstrate that ClusterGuard achieves over 2x efficiency improvement compared to advanced secure aggregation methods. Even with 20% of clients being malicious, the trained model maintains accuracy comparable to the original model, outperforming state-of-the-art robustness solutions. ClusterGuard provides a more efficient, secure, and robust solution for practical federated learning.