Optimized Tradeoffs for Private Prediction with Majority Ensembling

Authors: Shuli Jiang, Qiuyi Zhang, Gauri Joshi

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Lastly, we demonstrate the strong empirical effectiveness of our first-of-its-kind privacy-constrained utility optimization for ensembling labels for private prediction from private teachers in image classification. Notably, our Da RRM framework with an optimized γ exhibits substantial utility gains when compared against several baselines. Experiments. In downstream tasks, such as semi-supervised knowledge transfer for private image classification, we compare our Da RRM with an optimized γ to compute the private label majority from private teachers against PATE Papernot et al. (2018), which computes the private label majority from non-private teachers.
Researcher Affiliation Collaboration Shuli Jiang EMAIL Robotics Institute, Carnegie Mellon University Qiuyi (Richard) Zhang EMAIL Google Deep Mind Gauri Joshi EMAIL Electrical and Computer Engineering, Carnegie Mellon University
Pseudocode Yes Algorithm 1 Da RRM( ): Data-dependent Randomized Response Majority
Open Source Code Yes All code for the experiments can be found at https://anonymous.4open.science/r/Optimized Private Majority-CF50
Open Datasets Yes We use samples from two randomly chosen classes class 5 and 8 from the MNIST and Fashion-MNIST datasets to form our training and testing datasets.
Dataset Splits Yes Our MNIST has a total of 11272 training samples and 1866 testing samples; our Fashion-MNIST has 10000 training samples and 2000 testing samples. ... We train K = 11 teachers on equally divided subsets of the training datasets.
Hardware Specification No The paper mentions 'In practice, we observe with the Gurobi optimizer, one can optimize γ for K 41 on a laptop if δ > 0.' This refers to a practical limitation for optimization, not the specific hardware used for running the primary experiments or model training, and no specific model is mentioned.
Software Dependencies No The paper mentions 'Gurobi solver' and 'DP-SGD Abadi et al. (2016)' but does not provide specific version numbers for these software components or any other libraries or programming languages used.
Experiment Setup Yes Da RRM Setup: The Gaussian noise in DP-SGD has zero mean and std. σdpsgd = 12; the gradient norm clipping threshold is C = 1. ... We set the privacy allowance m = 35 ... We train K = 11 teachers ... for 5 epochs. ... We pick Q {20, 50, 100}.