Auditing Black-Box Prediction Models for Data Minimization Compliance
Authors: Bashir Rastegarpanah, Krishna Gummadi, Mark Crovella
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments with real-world prediction systems show that our auditing algorithms significantly outperform simpler benchmarks in both measurement and decision problems. Finally, we study the effectiveness of our auditing algorithms using different real-world prediction systems. We build prediction models by applying standard model evaluation and feature selection methods, and use the resulting models to perform black-box audits. Our experiments show that algorithms that exploit the proposed bandit framework significantly outperform simpler benchmarks. |
| Researcher Affiliation | Academia | Bashir Rastegarpanah Boston University bashir@bu.edu Krishna P. Gummadi MPI-SWS gummadi@mpi-sws.org Mark Crovella Boston University crovella@bu.edu |
| Pseudocode | Yes | Algorithm 1: Data Minimization Audit (Decision Problem) and Algorithm 2: Data Minimization Audit (Measurement Problem) are presented as structured pseudocode blocks. |
| Open Source Code | Yes | Code is available at: https://github.com/rastegarpanah/Data-Minimization-Auditor |
| Open Datasets | Yes | We build prediction systems using datasets from the UCI machine learning repository [7]. We use three datasets that have discrete features and discrete target variables... Digits/SVM. A dataset of 3823 images of hand-written digits from the MNIST [14] database is used. |
| Dataset Splits | No | 50% of data points are used to train the classifier (for Digits/SVM); A decision tree is built using 20% of data samples as the training data (for Census/Decision Tree and Nursery/Decision Tree). The paper specifies training set percentages but does not explicitly provide details for a separate validation split. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU or CPU models, processor types, or memory amounts used for running experiments. |
| Software Dependencies | No | The paper mentions using 'standard model evaluation and feature selection methods' and specific model types (e.g., 'support vector machine with linear kernel', 'decision tree') but does not specify any software libraries or frameworks with version numbers (e.g., Python, scikit-learn, PyTorch versions). |
| Experiment Setup | No | The paper mentions feature selection methods (e.g., 'recursive feature elimination procedure is applied to select 9 features') but does not provide specific experimental setup details such as hyperparameter values (e.g., learning rate, batch size, number of epochs), optimizer settings, or other system-level training configurations. |