Privacy-Preserving Federated Learning Based on Dataset Condensation

Federated Learning (FL) is a decentralized learning framework that facilitates learning from large-scale datasets while keeping raw data locally. It dismantles data silos and addresses privacy concerns tied to direct data access. However, the substantial interaction of model parameters between local nodes and servers in FL significantly impacts training efficiency. Moreover, inversion attacks on model weights may also lead to serious privacy leakage. Towards this end, we combine data condensation and differential privacy (DP) techniques, designing a communication-efficient and privacy-preserving FL scheme named PPFL-DC. Specifically, the designed DP-processing method protects the model weights trained on the real dataset. Then, the protected model weights are used to guide the generation of lightweight and privacy-preserving synthetic images, where an efficient synthetic data generation method with an appropriate number of iterations is designed to reduce local computation costs. Instead of large-sized model weights, these small synthetic images are sent to the server. Subsequently, the server recovers the model weights from the synthetic images and uses them for model aggregation to achieve highly accurate global model updates. PPFL-DC effectively reduces communication and computation overhead, ensures client-level privacy protection, and maintains a high global model accuracy, even when dealing with Non-IID datasets. Finally, the theoretical analysis and experimental results on three benchmark datasets demonstrate the superiority of PPFL-DC in performance over the previous FL schemes.