Research map / Federated learning and privacy / Differential privacy
Privacy Preserving Adaptive Experiment Design
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 DP-ConSE, a differentially private contextual bandit experiment algorithm for CATE estimation that trades off regret and estimation error, matching a lower bound (regret × MSE = O(M)) while being (ε, 1/n)-anticipating-private.
Problem
Adaptive experiments estimating conditional average treatment effects (CATE) must balance regret (patient welfare) against estimation accuracy, while protecting privacy of patient data under differential privacy.
Approach
Extends a regret-vs-ATE tradeoff framework to contextual bandits with per-feature successive elimination in the first half, randomized-length RCTs in the second half, and Laplace noise on rewards and batch lengths for anticipating differential privacy; shows privacy comes at no asymptotic cost in the regret-estimation Pareto curve.
Key results
- DP-ConSE is (ε, 1/n)-private and attains regret O(M max{f_min(n)^(1-α), log n/ε}) with estimation error O(1/max{f_min(n)^(1-α), log n/ε}).
- The product of regret and CATE MSE is O(M) under DP, matching the non-private lower bound Ω(M), so privacy is 'almost free' in the Pareto trade-off.
- DP-ConSE estimator is asymptotically normal with the same asymptotic variance as non-private ConSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ConSE to balance regret and CATE estimation error | Contextual bandit experiment · Contextual bandit experiment (CATE) | regret × estimation error: \mathcal{O}(M) versus lower bound \Omega(M) | Helped |
| Add differential privacy using DP-ConSE | ConSE · Contextual bandit experiment (CATE) | regret × estimation error: \mathcal{O}(M) | No clear effect |
| Vary α toward 1 (ablation) | ConSE · Contextual bandit experiment (CATE) | regret | Helped |
| Vary α toward 1 (ablation) | ConSE · Contextual bandit experiment (CATE) | estimation error (MSE) | Hurt |
| Vary α toward 0 (ablation) | ConSE · Contextual bandit experiment (CATE) | estimation error (MSE) | Helped |
| Vary α toward 0 (ablation) | ConSE · Contextual bandit experiment (CATE) | regret | Hurt |
| Add differential privacy | ConSE · Contextual bandit experiment (CATE) | asymptotic variance | No clear effect |
| Use UCB or Thompson sampling instead of an exploration-heavy experiment design | Contextual bandit experiment · Contextual bandit experiment (CATE) | estimation error (MSE): \Omega\left(\frac{M}{\log (n)}\right) | Hurt |
Surprising findings
- Privacy adds no asymptotic cost to the regret-estimation Pareto frontier; only the attainable minimum (log n/ε) shifts, shrinking the curve for stronger privacy.
Limitations
- Results are asymptotic and rely on seasonal non-stationarity and assumed constant CATE magnitude Δ=Θ(1).
- Analysis restricted to binary treatment arms and finite feature sets.
- Uses anticipating DP rather than classical DP, which protects only outputs after time t.
Experimental setup
Metrics: regret mean squared error of CATE estimate asymptotic normality
Baselines: ConSE (non-private) random controlled trials (RCT) UCB Thompson sampling
Builds on: ConSE (Bandit experiment regret-ATE tradeoff) anticipating differential privacy DP successive elimination tree mechanism
Closest papers by meaning
- Differentially private learners for heterogeneous treatment effects (ICLR 2025 · similarity 0.88)
- Shuffle and Joint Differential Privacy for Generalized Linear Contextual Bandits (NeurIPS 2026 · similarity 0.84)
- Improving the Variance of Differentially Private Randomized Experiments through Clustering (ICML 2025 · similarity 0.83)
- Towards Safe Policy Learning under Partial Identifiability: A Causal Approach (AAAI 2024 · similarity 0.83)
- Connecting Thompson Sampling and UCB: Towards More Efficient Trade-offs Between Privacy and Regret (ICML 2025 · similarity 0.83)
- Efficient Adaptive Experimentation with Noncompliance (NeurIPS 2025 · similarity 0.83)
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.