Researchers have developed a new method called Randomized Projection with Clipping (RPC) to defend against poisoning attacks in local differential privacy (LDP) protocols. This method is designed for users who possess multiple data items, a scenario where existing defenses are less effective due to larger output spaces that empower adversaries. RPC involves users projecting their local data onto a random vector sent by the server and then clipping the result to limit attack capabilities. The researchers also introduced a correction method to address clipping bias, allowing for a reduced clipping threshold and enhanced robustness without compromising the bias-variance tradeoff. Theoretical guarantees and numerical experiments demonstrate RPC's effectiveness in both trusted and untrusted environments, showing comparable or better performance than existing methods and significantly improved robustness against poisoning. AI
IMPACT Enhances privacy guarantees for distributed machine learning systems.
RANK_REASON Academic paper detailing a new method for privacy protocols. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
