Researchers have developed Cocoon, a new system architecture designed to improve the efficiency of differentially private machine learning training. This approach addresses the accuracy degradation typically associated with differential privacy by utilizing correlated noises, which cancel each other out over training iterations. Cocoon optimizes the handling of large noise histories across CPU, GPU, and specialized near-memory processing (NMP) devices, and includes specific optimizations for sparse embedding tables. Testing on an FPGA-based NMP prototype demonstrated performance improvements ranging from 1.23x to 10.82x. AI
IMPACT Enhances privacy-preserving ML training, potentially enabling wider adoption of sensitive data analysis.
RANK_REASON The cluster describes a research paper detailing a new system architecture for machine learning. [lever_c_demoted from research: ic=1 ai=1.0]
- central processing unit
- Cocoon
- differential privacy
- Donghwan Kim
- embedding tables
- field-programmable gate array
- graphics processing unit
- machine learning
- memory extension module
- near-memory processing (NMP)
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
