Reproduce “Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Roie Reshef, Kfir Yehuda Levy
Researchers at Technion developed a novel noise-cancellation mechanism within a federated learning framework, achieving optimal population loss guarantees with linear computational complexity (O(n)) for differentially private stochastic convex optimization. This method effectively supports partial-participation scenarios with an untrusted central server, demonstrating a significant improvement in efficiency compared to prior optimal approaches while maintaining comparable accuracy.
Key insights
- The inherent variance reduction property of μ²-SGD allows for strong differential privacy guarantees with minimal external noise, paving the way for optimal convergence rates.
- Injecting correlated noise across rounds, specifically `y_t,i - Y_{t-1,i}`, enables previous noise components to cancel out upon server aggregation, preventing unbounded noise accumulation.
- Adaptive noise calibration based on participation frequency (`N_t,i`) is crucial for maintaining privacy for each individual device while achieving optimal overall utility in partial-participation settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.