Researchers have introduced a new privacy framework called "privacy via predictability" that offers a more fine-grained approach than traditional differential privacy. This method accounts for an attacker's specific knowledge and the characteristics of the data, measuring privacy leakage by the attacker's ability to predict sensitive information. The framework is shown to be generally incomparable to differential privacy but can imply mutual-information differential privacy in certain worst-case scenarios. A new method using generalized method of moments (GMM) is proposed for analyzing predictability, leading to a predictability-calibrated output perturbation scheme for empirical risk minimization (ERM). AI
IMPACT Introduces a novel privacy metric that could lead to more efficient privacy-accuracy trade-offs in AI models.
RANK_REASON The cluster contains an academic paper detailing a new theoretical framework for privacy in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]
- Differential Privacy
- Empirical Risk Minimization
- Generalized Method of Moments
- Privacy via predictability
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
