Differentially-Private Software Frequency Profiling under Linear Constraints
summary
Reformulates local differential privacy for frequency vectors that must satisfy software-specific linear inequalities and gives a concrete noise calibration: Laplace noise scale ≥ 2 D_v/ε, where D_v is computed by solving an integer linear program (and for pairwise constraints by constraint graph reachability). Implements an opt-in sample to bound τ and uses scale 2τ/ε for all users. Evaluated on 15 Android apps with 1,000 simulated users, achieving average normalized error <0.07 when protecting half or a quarter of methods at ε=0.5.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:24Remote software profiling
- 1:56Example: method frequency profiling
- 2:16Data privacy
- 4:53One choice for R is to add Laplace noise
- 5:44Protection goal: hiding presence
- 7:45Linear constraints on frequencies
- 8:44Example: hiding b's presence
- 13:29Contributions
related talks
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
01h 00mDifferential Privacy: Fundamentals to Forefront I
01h 03mStatistical Inference and Privacy, Part II
24mLocally Differentially Private Protocols for Frequency Estimation
58mLarge-Scale Private Learning, Part III
47mFinite Sample Differentially Private Confidence Intervals
42mDifferential Privacy
42mDifferential Privacy: From Theory to Deployment