#gaussian-mechanisms News & Analysis

2 articles tagged with #gaussian-mechanisms. AI-curated summaries with sentiment analysis and key takeaways from 50+ sources.

2 articles

AINeutralarXiv – CS AI · May 286/10

🧠

Mind the Gap: Mixtures of Gaussians in Approximate Differential Privacy

Researchers introduce mixture mechanisms for differential privacy that combine multiple Gaussian distributions to reduce noise in data queries while maintaining privacy guarantees. These mechanisms substantially outperform existing analytic Gaussian approaches in low-privacy regimes, approaching theoretical optimality with significantly lower noise amplitudes and variances.

AINeutralarXiv – CS AI · May 276/10

🧠

Hidden-State Privacy Has an Empty Middle

Researchers demonstrate that Gaussian mechanisms for hidden-state privacy face a fundamental trade-off, with no configurations achieving both moderate utility and moderate privacy against adaptive attackers. A diagonal inverse-Fisher mechanism emerges as minimax-optimal but sits at the privacy-utility boundary rather than within an achievable middle ground, suggesting future work must redesign architectures rather than optimize within existing Gaussian frameworks.