Research map / Federated learning and privacy / Differential privacy
Protect Your Score: Contact-Tracing with Differential Privacy Guarantees
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DPFN, a differentially private factorized-neighbors algorithm that releases contact-tracing risk scores with log-normal noise on message products, evaluated on OpenABM and Covasim simulators.
Problem
Contact tracing risk scores, when released to users, let an adversary infer a victim's private health status; existing approaches lack differential privacy guarantees for this release while keeping utility for pandemic mitigation.
Approach
Decompose the factorized-neighbors update into products of contact messages and add log-normal noise to each product, whose closed-form Rényi divergence yields an (ε, δ)-DP guarantee for the released COVIDSCORE, with noise introduced later in the computation than per-message noising.
Key results
- At epsilon=1 per message, DPFN achieves peak infection rate below 1% on simulators, while other DP methods need about ten times larger epsilon to match.
- DPFN achieves two to ten-fold lower peak infection rate than traditional contact tracing at epsilon=1; e.g., 1.5h vs 121.6h at 100,000 agents.
- At 1 million agents, DPFN gives 0.2h peak infection rate versus 134.4h for traditional contact tracing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DPFN with ε = 1 per message instead of traditional contact tracing | Contact tracing · OpenABM and Covasim | peak infection rate | Helped |
| use DPFN instead of traditional contact tracing | Contact tracing · OpenABM (50,000 agents) | peak infection rate (h): 123.2 -> 2.5 | Helped |
| use DPFN instead of traditional contact tracing | Contact tracing · OpenABM (100,000 agents) | peak infection rate (h): 121.6 -> 1.5 | Helped |
| use DPFN instead of traditional contact tracing | Contact tracing · OpenABM (500,000 agents) | peak infection rate (h): 134.9 -> 0.5 | Helped |
| use DPFN instead of traditional contact tracing | Contact tracing · OpenABM (1,000,000 agents) | peak infection rate (h): 134.4 -> 0.2 | Helped |
| apply ε = 1 differential privacy to FN | Factorised Neighbors · OpenABM with noisy tests (FPR 25%; FNR 3%) | peak infection rate (h): 0.6 -> 46.6 | Hurt |
| increase daily testing budget from 10% to 15% | DPFN · OpenABM with noisy tests (FPR 25%; FNR 3%) | peak infection rate (h): 46.6 -> 0.7 | Helped |
| use DPFN instead of per-message FN | Contact tracing · OpenABM and Covasim at ε = 1 | peak infection rate | Helped |
| use DPFN instead of Gibbs sampling | Contact tracing · OpenABM and Covasim | peak infection rate | Helped |
Surprising findings
- Gibbs sampling with DP performs worse than all other methods in peak infection rate, despite being the strongest pure-ε-DP option.
Limitations
- Assumes full adoption of the contact tracing app and does not address partial adoption.
- Privacy bound degrades with repeated contacts with an adversary; advanced composition not yet analyzed.
- Noisy tests (high FPR) make DPFN much less robust unless more tests are performed.
Experimental setup
Datasets: OpenABM simulator Covasim simulator
Benchmarks: OpenABM COVID19 simulator Covasim simulator
Metrics: peak infection rate privacy epsilon (differential privacy)
Baselines: Traditional contact tracing Gibbs sampling with DP Per-message FN
Builds on: Factorized Neighbors Gaussian mechanism Rényi differential privacy Gibbs sampling DP (Wang et al. 2015)
Closest papers by meaning
- Enhancing Privacy in the Early Detection of Sexual Predators Through Federated Learning and Differential Privacy (AAAI 2025 · similarity 0.81)
- Controlling The Spread of Epidemics on Networks with Differential Privacy (NeurIPS 2025 · similarity 0.81)
- Privacy Guarantees in Posterior Sampling under Contamination (NeurIPS 2026 · similarity 0.79)
- Privacy amplification by random allocation (NeurIPS 2025 · similarity 0.78)
- Differentially Private Worst-group Risk Minimization (ICML 2024 · similarity 0.77)
- Optimal Best Arm Identification under Differential Privacy (NeurIPS 2025 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.