Research map / Federated learning and privacy / Differential privacy
Instance-Wise Laplace Mechanism via Deep Reinforcement Learning (Student Abstract)
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 an instance-wise Laplace mechanism that assigns non-identical Laplace noise per data instance, with the noise variances chosen by a DQN agent so the output satisfies epsilon-pDP while preserving data statistics better than the standard Laplace mechanism.
Problem
Standard Laplace mechanisms add identical noise to all query outputs, which distorts data statistics, while per-instance differential privacy (pDP) implies each instance has different privacy loss; the interaction of noise across instances makes optimizing non-identical noise hard.
Approach
Formulate the choice of per-instance Laplace noise variances as a Markov decision process and solve it with a DQN-based deep reinforcement learning agent, with a reward that enforces epsilon-pDP and minimizes KL-divergence to the original query output distribution.
Key results
- Proposed instance-wise Laplace mechanism achieves KL-divergence 0.883 versus 1.968 for the conventional Laplace mechanism at epsilon = 0.011 on the height data.
- The method assigns lower-variance noise to frequently occurring data instances, consistent with lower privacy risk for them.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use instance-wise Laplace noise instead of conventional Laplace noise | Laplace mechanism · Personal height dataset | KL-divergence: 1.968 -> 0.883 | Helped |
Limitations
- Method lacks generality across noise distributions such as Gaussian.
- Evaluated only on a single real dataset with a single small epsilon value.
- Small-scale student abstract; details of DQN training and hyperparameters are not reported.
Experimental setup
Datasets: NBA players height dataset (Kaggle)
Benchmarks: NBA players height dataset (Kaggle)
Metrics: KL divergence noise standard deviation
Baselines: Laplace mechanism
Builds on: Laplace mechanism per-instance differential privacy DQN
Closest papers by meaning
- Beyond the Calibration Point: Mechanism Comparison in Differential Privacy (ICML 2024 · similarity 0.87)
- Private Hyperparameter Tuning with Ex-Post Guarantee (NeurIPS 2025 · similarity 0.87)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.87)
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · similarity 0.86)
- Leveraging Per-Instance Privacy for Machine Unlearning (ICML 2025 · similarity 0.86)
- Rényi Pufferfish Privacy: General Additive Noise Mechanisms and Privacy Amplification by Iteration via Shift Reduction Lemmas (ICML 2024 · similarity 0.86)
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.