Multi-Task Differential Privacy Under Distribution Skew
Authors: Walid Krichene, Prateek Jain, Shuang Song, Mukund Sundararajan, Abhradeep Guha Thakurta, Li Zhang
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental studies on recommendation problems that exhibit a long tail of small tasks, demonstrate that our methods significantly improve utility, achieving the state of the art on two standard benchmarks. We conduct experiments on synthetic data and two standard recommendation benchmarks which exhibit a heavy distribution skew. |
| Researcher Affiliation | Industry | 1Google Research 2Google. Correspondence to: Walid Krichene <walidk@google.com>. |
| Pseudocode | Yes | Algorithm 1 Task-weighted SSP (Sufficient Statistic Perturbation); Algorithm 2 Task-weighted noisy gradient descent; Algorithm 3 Alternating Minimization for Private Matrix Completion |
| Open Source Code | Yes | The code is available at the following repository: https://github.com/google-research/google-research/tree/master/ dp alternating minimization. The code used to run the experiments is made available at the following URL: https://github.com/google-research/google-research/tree/master/dp alternating minimization. |
| Open Datasets | Yes | Experimental studies on recommendation problems that exhibit a long tail of small tasks, demonstrate that our methods significantly improve utility, achieving the state of the art on two standard benchmarks. We conduct experiments on synthetic data and two standard recommendation benchmarks which exhibit a heavy distribution skew. Movie Lens data sets (Harper & Konstan, 2016). Million Song Data (Bertin-Mahieux et al., 2011) data sets. |
| Dataset Splits | No | The paper states: 'We partition the data into training and test sets following an 80-20 random split.' for synthetic data, and 'We follow the same experimental setting as (Jain et al., 2018; Chien et al., 2021; Jain et al., 2021)' for real data. While this indicates how data is split for training and testing, it does not explicitly specify a separate validation set percentage or sample count, nor does it refer to pre-defined validation splits. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running the experiments. It does not mention GPU models, CPU types, memory specifications, or cloud computing instance types. |
| Software Dependencies | No | The paper refers to using standard algorithms like DPSGD and DPALS, and mentions the use of 'the source code' but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, specific libraries, or compilers with their versions). |
| Experiment Setup | Yes | When computing adaptive weights for our methods, we first compute private estimates ˆni of the counts...then define task weights following eq. (11)-(12), but allowing a wider range of exponents. More precisely, we replace eq. (11) with ω i = ˆn µ i q Pm i =1 ˆn 2µ+1 i /n β , (16) where µ is a hyper-parameter. We set δ = 10 5 for ML10M and δ = 1/n for ML20M and MSD. The proportion of RDP budget spent on estimating counts (out of the total RDP budget) is 20% for ϵ = 20, 14% for ϵ = 5 and 12% for ϵ = 1. This was tuned on the DPALS baseline (with tail sampling). |