Revisiting Gradient Sparsity Assumptions in Differentially Private Federated Learning

Key Points

• This research aims to challenge traditional assumptions related to gradient sparsity in differentially private federated learning.
• Conducted empirical experiments using multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs).
• Analyzed the relationship between gradient sparsity and model performance during training.
• Assessed the impact of architecture differences on sparsity.
• MLPs exhibited higher gradient sparsity compared to CNNs.
• Observed that gradient sparsity decreases as training progresses.
• Found that higher sparsity does not correlate with improved performance.