A Huber Loss Minimization Approach to Mean Estimation under User-level Differential Privacy
Authors: Puning Zhao, Lifeng LAI, Li Shen, Qingming Li, Jiafei Wu, Zhe Liu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we perform numerical experiments to validate our theoretical analysis. In this section, we show numerical experiments. We compare the performance of our new Huber loss minimization approach (denoted as HLM) versus the two-stage approach proposed in [34], called Winsorized Mean Estimator (denoted as WME). |
| Researcher Affiliation | Collaboration | Puning Zhao Zhejiang Lab pnzhao@zhejianglab.com Lifeng Lai University of California, Davis lflai@ucdavis.edu Li Shen Sun Yat-Sen University mathshenli@gmail.com Qingming Li Zhejiang University liqm@zju.edu.cn Jiafei Wu Zhejiang Lab wujiafei@zhejianglab.com Zhe Liu Zhejiang Lab zhe.liu@zhejianglab.com |
| Pseudocode | Yes | The update rule is Pn i=1 wi min n 1, Ti ck yi o yi Pn i=1 wi min n 1, Ti ck yi o . (23) The algorithm can be designed from above equation. Suppose that the algorithm starts from c0. The update rule is... (23) is run iteratively until the norm of update ck+1 ck between two iterations is less than ξ. |
| Open Source Code | Yes | Answer: [Yes] Justification: Codes are provided. |
| Open Datasets | Yes | Finally, Figure 1 (g) and (h) show the results using the IPUMS dataset [83] for total income and salary, respectively. Ruggles, S., S. Flood, M. Sobek, et al. IPUMS USA: Version 15.0 [dataset], 2024. |
| Dataset Splits | No | The paper mentions tuning parameters optimally for each case, but does not explicitly provide training, validation, or test dataset splits. |
| Hardware Specification | No | Answer: [NA] Justification: The experiments need little computational resources. |
| Software Dependencies | No | The paper mentions an algorithm modification, but does not provide specific software names with version numbers. |
| Experiment Setup | Yes | In the following experiments, we fix ϵ = 1 and δ = 10 5. For a fair comparison, the parameter T for our method as well as τ in [34] are both tuned optimally for each case. In each curve, n is fixed, while the number of samples per user m varies from 1 to 1, 000. |