Researchers have developed a new method for optimizing privacy allocation in machine learning, specifically for Differentially Private Stochastic Gradient Descent (DPSGD). The approach formulates the Gaussian noise-variance schedule as a control problem under a Concentrated Differential Privacy constraint. This work introduces Privacy-Budget Adjoint Mirror Descent (PB-AMD), which combines adjoint gradients with entropic updates to ensure positivity and budget adherence. The method aims to improve training utility by optimizing privacy spending across iterations, demonstrating better performance than existing allocation rules and faster optimization times compared to projected gradient descent. AI
IMPACT This research could lead to more efficient and effective privacy-preserving machine learning models by optimizing how privacy budgets are utilized during training.
RANK_REASON Academic paper detailing a new method for privacy allocation in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Concentrated Differential Privacy
- Differentially Private Stochastic Gradient Descent
- Gaussian function
- Kalman filter
- PB-AMD
- Privacy-Budget Adjoint Mirror Descent
- Projected Gradient Descent
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
