Khái niệm cốt lõi
SPriFed-OMP algorithm ensures accurate sparse basis recovery in FL with DP.
Tóm tắt
The article introduces SPriFed-OMP, a novel algorithm for sparse basis recovery in Federated Learning (FL) with Differential Privacy (DP). It addresses the challenge of protecting client data privacy while accurately recovering sparse models. The algorithm combines Orthogonal Matching Pursuit (OMP) with secure multi-party computation and DP to achieve efficient and accurate sparse basis recovery. The article provides theoretical analysis and empirical results demonstrating the algorithm's superior performance compared to existing DP-FL solutions.
- Introduction
- Sparse basis recovery is crucial in statistical learning applications with limited samples.
- Federated Learning (FL) protects user privacy by keeping data on clients during model training.
- Data Privacy in FL
- FL updates models on client data without uploading data to the server.
- Differential Privacy (DP) is introduced to protect client gradients in FL.
- Challenges in Sparse Recovery
- Existing DP-FL algorithms struggle with accurate sparse recovery in high-dimensional settings.
- SPriFed-OMP algorithm overcomes the curse-of-dimensionality in DP-FL for sparse recovery.
- Algorithm Design
- SPriFed-OMP combines OMP with DP and secure computation for efficient sparse basis recovery.
- Enhanced version SPriFed-OMP-GRAD improves performance with gradient privatization.
- Privacy Analysis
- Algorithms ensure privacy with Gaussian Differential Privacy (GDP) and L2-sensitivity bounds.
- Accuracy Guarantee
- Theorems 7 and 8 provide conditions for SPriFed-OMP and SPriFed-OMP-GRAD to recover true sparse basis with high probability.
Thống kê
"In particular, for DP-SGD (Abadi et al., 2016), even when it is applied to Lipschitz loss functions, the empirical risk is of the order Op p nq (Theorem 2.4 in Bassily et al. (2014))."
"Other DP-FL algorithms, such as objective perturbation (Kifer et al., 2012), have similar issues."
Trích dẫn
"SPriFed-OMP algorithm ensures accurate sparse basis recovery in FL with DP."