Privacy Amplification for Matrix Mechanisms
Authors: Christopher A. Choquette-Choo, Arun Ganesh, Thomas Steinke, Abhradeep Guha Thakurta
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our algorithm also has practical empirical utility. We show that amplification leads to significant improvement in the privacy/utility trade-offs for DP-FTRL style algorithms for standard benchmark tasks. ... 6 EMPIRICAL IMPROVEMENTS We implement MMCC by building on methods in the open-source dp accounting Python library (DP Team, 2022), and perform empirical studies of the amplification benefits from MMCC. |
| Researcher Affiliation | Industry | Christopher A. Choquette-Choo Arun Ganesh Thomas Steinke Abhradeep Thakurta Google Deep Mind. cchoquette@google.com. Google Research. arunganesh@google.com. Google Deep Mind. steinke@google.com. Google Deep Mind. athakurta@google.com. |
| Pseudocode | Yes | Algorithm 1 Matrix Mechanism Conditional Composition algorithm, MMCC(C, p, σ, δ1, δ2) |
| Open Source Code | Yes | We plan to publicly release a library implementing MMCC with the final manuscript. ... Our implementation is currently open-sourced as part of the dp accounting library, and PLD accounting for Mo G mechanisms can be done using dp accounting.pld.PLDAccountant and dp accounting.dp event.Mixture Of Gaussians Dp Event. ... https://github.com/google/differential-privacy/tree/main/python/dp_ accounting/dp_accounting |
| Open Datasets | Yes | We reproduce the centralized DP training on CIFAR-10 from Choquette-Choo et al. (2023b), including model architecture, tuning setup, hyperparameter choices, and optimizations to the tree aggregation mechanism for ML; we use these as our baseline results. |
| Dataset Splits | No | The paper mentions using 'standard benchmark tasks' and reproducing 'centralized DP training on CIFAR-10 from Choquette-Choo et al. (2023b)' but does not provide specific details on the training, validation, or test dataset splits used for reproducibility. |
| Hardware Specification | No | The paper does not provide specific details regarding the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'the open-source dp accounting Python library (DP Team, 2022)' and 'dp accounting.pld.PLDAccountant and dp accounting.dp event.Mixture Of Gaussians Dp Event' but does not provide specific version numbers for these software components or for Python itself, which is necessary for reproducibility. |
| Experiment Setup | No | The paper states, 'We reproduce the centralized DP training on CIFAR-10 from Choquette-Choo et al. (2023b), including model architecture, tuning setup, hyperparameter choices, and optimizations to the tree aggregation mechanism for ML,' indicating these details are in another source, but they are not explicitly provided within this paper. |