Are We There Yet? Timing and Floating-Point Attacks on Differential Privacy Systems

Explore timing and floating-point attacks on differential privacy systems in this IEEE conference talk. Delve into data analysis privacy leakage, differential privacy concepts, and additive noise techniques. Examine threat models, floating-point implementations, and attacks on Gaussian mechanisms. Investigate discrete distributions, timing side-channels, and positive correlations in timing attacks. Learn about mitigation strategies for both floating-point and timing attacks, gaining insights into the vulnerabilities and defenses of privacy-preserving systems.

Intro
Contributions
Data Analysis: Privacy Leakage
Data Analysis: Differential Privacy
DP: Additive Noise
DP: Deployments and Libraries
Threat model
FP Attack: FP Implementation
FP Attack: Impossible Outputs
FP Attack: Gaussian Implementation
FP Attack: IsFeasible(s)
FP Attack: Attack Results
FP Attack: DP-SGD
Discrete Distributions: Implementation
Discrete Distributions: Timing Side-Channel
Timing Attack: Positive Correlation
Timing Attack: Attack Results
Mitigation: FP attack
Mitigation: Timing attack