Research map / Federated learning and privacy / Privacy attacks and secure learning
Frequency Oracle for Sensitive Data Monitoring (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.
Empirically benchmarks local differential privacy frequency oracle algorithms (e.g., Direct/Unary/Histogram Encoding, RAPPOR, Hadamard, Fast Local Hashing) across data size, peak, epsilon, and federated learning settings on smart-meter energy data to recommend which to use.
Problem
Choosing the appropriate local differential privacy frequency oracle algorithm for a given data type and setting is unclear; practitioners lack guidance on which FO to use when.
Approach
Organizes frequency oracles into a taxonomy (heavy hitters, sketches, encoding algorithms with segment-pair and Hadamard groups) and empirically compares them under varying data properties, privacy budget, and a federated learning setting where model weights are privatized.
Key results
- Direct Encoding gives highest accuracy for small datasets, Unary Encoding for large datasets, and RAPPOR for massive datasets.
- In federated learning, Unary Encoding, Histogram Encoding, Fast Local Hashing, and Hadamard Response remained accurate under privatization, while Count Mean Sketch and Hadamard Mechanism struggled.
- Privatizing model weights showed smaller error than privatizing data on the client side.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Direct Encoding for small datasets | Frequency Oracle algorithms · Frequency estimation on small datasets | accuracy | Helped |
| Use Unary Encoding for large datasets | Frequency Oracle algorithms · Frequency estimation on large datasets | accuracy | Helped |
| Use RAPPOR for massive datasets | Frequency Oracle algorithms · Frequency estimation on massive datasets | accuracy | Helped |
| Increase the peak in the data | Frequency Oracle algorithms · Frequency estimation | accuracy | No clear effect |
| Increase the privacy parameter ϵ | Frequency Oracle algorithms · Frequency estimation | accuracy | Helped |
| Privatize model weights instead of client-side data | Federated learning · Federated learning | error | Helped |
| Use Unary Encoding, Histogram Encoding, Fast Local Hashing, or Hadamard Response with privatized data | Frequency Oracle algorithms · Federated learning | prediction accuracy | Helped |
| Use Count Mean Sketch or Hadamard Mechanism with privatized data | Frequency Oracle algorithms · Federated learning | prediction accuracy | Hurt |
| Use a high rather than low ϵ | Direct Encoding and Sketch Response · Federated learning | prediction accuracy | Helped |
| Change ϵ | RAPPOR · Federated learning | prediction accuracy | No clear effect |
Surprising findings
- Changes in the range of data values (peaks) had a negligible effect on accuracy.
Limitations
- Preliminary study with only two energy datasets; authors note need for bigger and more diverse datasets.
- Federated learning evaluation limited to one configuration shown (Unary Encoding with epsilon=3).
Experimental setup
Datasets: Smart meters in London (UKPowerNetworks 2018) Energy consumption in Germany (Milojkovic 2021)
Benchmarks: London smart meter data German energy consumption data
Metrics: frequency estimation accuracy prediction error
Baselines: Direct Encoding Unary Encoding Histogram Encoding RAPPOR Fast Local Hashing Hadamard Response Count Mean Sketch Hadamard Mechanism Sketch Response
Builds on: Frequency oracles (Cormode, Maddock, and Maple 2021)
Closest papers by meaning
- Private and Personalized Frequency Estimation in a Federated Setting (NeurIPS 2024 · similarity 0.83)
- Oracle-Efficient Differentially Private Learning with Public Data (NeurIPS 2024 · similarity 0.81)
- Noise-Aware Algorithm for Heterogeneous Differentially Private Federated Learning (ICML 2024 · similarity 0.80)
- Monitoring Violations of Differential Privacy over Time (NeurIPS 2026 · similarity 0.79)
- Auditing $f$-differential privacy in one run (ICML 2025 · similarity 0.79)
- One-shot Empirical Privacy Estimation for Federated Learning (ICLR 2024 · similarity 0.79)
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.