Wasserstein Distance-Based Deep Leakage from Gradients.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) protects data privacy by sharing average gradients.
• Gradient-based attacks, like Deep Leakage from Gradient (DLG), can reconstruct private training data from shared gradients.
• Existing DLG methods suffer from slow convergence and poor reconstruction accuracy.

Purpose of the Study:

• To introduce a novel Wasserstein distance-based DLG (WDLG) method to enhance privacy protection in federated learning.
• To improve the convergence speed and accuracy of gradient-based privacy attacks.
• To address the limitations of existing DLG algorithms.

Main Methods:

• Proposed WDLG method utilizing Wasserstein distance as the training loss function.
• Iterative calculation of Wasserstein distance using Lipschitz condition and Kantorovich-Rubinstein duality.
• Theoretical analysis to prove the differentiability and continuity of Wasserstein distance.

Main Results:

• WDLG demonstrates superior performance over DLG in terms of training speed and inversion image quality.
• Experimental validation confirms the effectiveness of WDLG in enhancing gradient-based attack reconstruction.
• Differential privacy is shown to provide effective disturbance protection against WDLG.

Conclusions:

• The WDLG method offers a significant improvement for gradient-based privacy attacks in federated learning.
• Wasserstein distance is a viable and effective loss function for improving attack accuracy and convergence.
• Differential privacy integration presents a promising direction for robust privacy-preserving deep learning frameworks.