Researchers have developed new methods to enhance differential privacy in machine learning, particularly for decentralized and causal structure learning. One approach, DPDL, uses a similarity-based calibration technique with Gaussian noise to protect privacy in decentralized learning on non-IID data, achieving linear speedup. Another method leverages fully homomorphic encryption (FHE) to enable calculations on encrypted data for causal structure learning, employing circuit simplification and approximations to manage FHE's computational cost. Additionally, a new DP sketching mechanism based on fast transforms offers improved runtime for DP linear regression, and a local differential privacy model with correlated noise demonstrates that optimal centralized privacy costs can be achieved without sacrificing utility. AI
IMPACT These advancements in differential privacy techniques could enable more secure and robust AI systems, particularly in decentralized learning and causal inference.
RANK_REASON Multiple academic papers published on arXiv detailing novel methods for differential privacy in machine learning.
- arXiv
- Centralized differential privacy
- Differential Privacy
- DP linear regression
- DP ordinary least squares
- Gaussian sketching
- Hadamard matrix
- Local differential privacy
- DPDL
- Fully Homomorphic Encryption
- Gaussian noise
- Gaussian sketch
- Newton-Raphson Reciprocal
- SIMD-acceleration
- Taylor expansion
AI-generated summary · Google Gemini · from 5 sources. How we write summaries →
