Model Reconstruction Using Counterfactual Explanations: A Perspective From Polytope Theory

Authors: Pasan Dissanayake, Sanghamitra Dutta

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our strategy improves fidelity between the target and surrogate model predictions on several datasets. and We conduct experiments on both synthetic datasets, as well as, four real-world datasets, namely, Adult Income [Becker and Kohavi, 1996], COMPAS [Angwin et al., 2016], DCCC [Yeh, 2016], and HELOC [FICO, 2018].
Researcher Affiliation Academia Pasan Dissanayake University of Maryland College Park, MD pasand@umd.edu Sanghamitra Dutta University of Maryland College Park, MD sanghamd@umd.edu
Pseudocode Yes Algorithm 1 Counterfactual Clamping
Open Source Code Yes A python-based implementation is available at: https://github.com/pasandissanayake/ model-reconstruction-using-counterfactuals.
Open Datasets Yes We use four publicly available real-world tabular datasets (namely, Adult Income, COMPAS, DCCC, and HELOC) to evaluate the performance of the attack proposed in Section 3.3. and Adult Income: The dataset is a 1994 census database with information such as educational level, marital status, age and annual income of individuals [Becker and Kohavi, 1996].
Dataset Splits No No explicit training/validation/test split percentages or sample counts for a separate validation set are provided for the primary model training or evaluation, beyond mentioning 'test-train-attack split' and using 'Dref' for fidelity evaluation.
Hardware Specification Yes All the experiments were carried-out on two computers, one with a NVIDIA RTX A4500 GPU and the other with a NVIDIA RTX 3050 Mobile.
Software Dependencies No The paper mentions a 'python-based implementation' but does not specify particular software versions for libraries, frameworks, or programming languages.
Experiment Setup Yes The hidden layer activations are Re LU and the layer weights are L2 regularized. The regularization coefficient is 0.001. Each model is trained for 200 epochs, with a batch size of 32.