Interpretable and Differentially Private Predictions

Authors: Frederik Harder, Matthias Bauer, Mijung Park4083-4090

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the usefulness of our approach on several image benchmark datasets as well as a medical dataset.In this section we evaluate the trade-off between accuracy, privacy, and interpretability for our LLM model on several datasets and compare to other methods where it is appropriate.
Researcher Affiliation Academia Frederik Harder,1,2 Matthias Bauer,1,3* Mijung Park1,2 1Max Planck Institute for Intelligent Systems, T ubingen, Germany 2Department of Computer Science, University of T ubingen, T ubingen, Germany 3Department of Engineering, University of Cambridge, Cambridge, UK
Pseudocode Yes Algorithm 1 DP-LLM for interpretable classification
Open Source Code Yes Our implementation is available on Git Hub2. 2github.com/frhrdr/dp-llm
Open Datasets Yes MNIST (Le Cun and Cortes 2010) and Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017) imagesand Henan Renmin Hospital Data (Li et al. 2017; Maxwell et al. 2017)4. It contains 110,300 medical records with 62 input features and 3 binary outputs.
Dataset Splits No We consider the classification of MNIST (Le Cun and Cortes 2010) and Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017) images with the usual train/test splits and train a CNN3 as a baseline model...The paper mentions train/test splits but does not specify a validation split explicitly or a detailed methodology for creating all splits (e.g., percentages or counts).
Hardware Specification No No specific hardware details (e.g., CPU, GPU models, or specific cloud instances) used for running the experiments were provided.
Software Dependencies No No specific software dependencies with version numbers (e.g., PyTorch 1.9, Python 3.8) were explicitly mentioned.
Experiment Setup Yes We train several LLMs in the private and nonprivate setting. By default, we use LLM models with M = 30 filters per class and random projections to D = 300 dimensions, which are optimized for 20 epochs using the Adam optimizer with learning rate 0.001, decreasing by 20% every 5 epochs. On MNIST the model benefits from a decreased inverse softmax temperature β = 1/30, while β = 1 is optimal for Fashion-MNIST. We choose a large batch size of 500, as this improves the signal-to-noise ratio of our algorithm. In the private setting we clip the persample gradient norm to C = 0.001 and train with σ = 1.3and By default, we use an LLM model with M = 2 filters per class and no random projections, which is optimized for 20 epochs using the Adam optimizer with learning rate 0.01, decreasing by 20% every 5 epochs. We choose a batch size of 256. In the private setting we clip the per-sample gradient norm to 0.001 and train with σ = 1.25.