Privacy-preserving clustered federated learning via differential privacy and homomorphically encrypted prototypes

Area of Science:

• Machine Learning
• Cryptography
• Distributed Systems

Background:

• Clustered Federated Learning (CFL) addresses statistical heterogeneity by grouping clients.
• Existing CFL methods risk exposing sensitive clustering information and updates, compromising privacy.
• There is a need for privacy-preserving CFL methods that protect both data and clustering signals.

Purpose of the Study:

• To propose a novel Privacy-Preserving Clustered Federated Learning (PPCFL) framework.
• To enhance privacy protection for clustering signals and cluster-specific updates in CFL.
• To improve the privacy-utility trade-off in dynamic Non-IID federated learning settings.

Main Methods:

• A split-stream framework integrating adaptive Gaussian perturbation and threshold Paillier encrypted aggregation.
• Backbone updates are protected by adaptive Gaussian perturbation; clustering signatures and cluster-head updates use stream-specific perturbations and Paillier encryption.
• Server aggregates in ciphertext domain; decryption is performed by a qualified client subset, preventing server decryption.

Main Results:

• PPCFL achieves the highest final-round accuracy across MNIST, Fashion-MNIST, and CIFAR-10 datasets compared to existing methods.
• Demonstrates enhanced protection for clustering-related information and cluster-specific updates.
• Improves final accuracy over DP-FedAvg and IFCA under Dirichlet settings, showing significant gains on various datasets.

Conclusions:

• PPCFL effectively enhances privacy in Clustered Federated Learning without sacrificing model performance.
• The proposed framework offers a robust solution for privacy-preserving collaborative machine learning under heterogeneous data distributions.
• PPCFL represents a significant advancement in secure and efficient federated learning systems.