Fooling SHAP with Stealthily Biased Sampling

Authors: gabriel laberge, Ulrich Aïvodji, Satoshi Hara, Mario Marchand, Foutse Khomh

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally (Section 5), we illustrate the impact of the proposed manipulation attack on a synthetic dataset and four popular datasets, namely Adult Income, COMPAS, Marketing, and Communities. We observed that the proposed attack can reduce the importance of a sensitive feature while keeping the data manipulation undetected by the audit.
Researcher Affiliation Academia 1Polytechnique Montréal, Québec 2École de technologie supérieure, Québec 3Osaka University, Japan 4Universitié de Laval à Québec
Pseudocode Yes Algorithm 1 Compute non-uniform weights
Open Source Code Yes The source code of all our experiments is available online3.
Open Datasets Yes We consider four standard datasets from the FAcc T literature, namely COMPAS, Adult-Income, Marketing, and Communities.
Dataset Splits Yes The datasets were first divided into train/test subsets with ratio 4:5. The models were trained on the training set and evaluated on the test set. All categorical features for COMPAS, Adult, and Marketing were one-hot-encoded which resulted in a total of 11, 40, and 61 columns for each dataset respectively. A simple 50 steps random search was conducted to fine-tune the hyper-parameters with cross-validation on the training set.
Hardware Specification No No specific hardware details (like GPU models, CPU types, or memory) used for running experiments are mentioned in the paper.
Software Dependencies No The paper mentions the 'SHAP Python library' and that some parts were rewritten in 'C++', but no specific version numbers for these or other software dependencies are provided.
Experiment Setup Yes Three models were considered for the two datasets: Multi-Layered Perceptrons (MLP), Random Forests (RF), and e Xtreme Gradient Boosted trees (XGB). One model of each type was fitted on each dataset for 5 different train/test splits seeds, resulting in 60 models total. A simple 50 steps random search was conducted to fine-tune the hyper-parameters with cross-validation on the training set.