Auditing Local Explanations is Hard

Authors: Robi Bhattacharjee, Ulrike Luxburg

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

Reproducibility Variable Result LLM Response
Research Type Theoretical We prove upper and lower bounds on the amount of queries that are needed for an auditor to succeed within this framework. Our results show that successful auditing requires a potentially exorbitant number of queries particularly in high dimensional cases. Our analysis also reveals that a key property is the locality of the provided explanations a quantity that so far has not been paid much attention to in the explainability literature.
Researcher Affiliation Academia Robi Bhattacharjee University of Tübingen and Tübingen AI Center robi.bhattacharjee@wsii.uni-tuebingen.de Ulrike von Luxburg University of Tübingen and Tübingen AI Center ulrike.luxburg@uni-tuebingen.de
Pseudocode Yes Algorithm 1 simple_audit(X, f(X), E(f, X), ϵ1, ϵ2, γ, δ)
Open Source Code No This is a theory paper, and therefore does not provide open-source code for a practical methodology.
Open Datasets No This is a theory paper and does not conduct experiments with real-world datasets. It discusses theoretical data distributions like 'µ'.
Dataset Splits No This is a theory paper and does not conduct experiments with data splits.
Hardware Specification No This is a theory paper and does not conduct experiments requiring specific hardware specifications.
Software Dependencies No This is a theory paper and does not conduct experiments requiring specific software dependencies with version numbers.
Experiment Setup No This is a theory paper and does not describe an experimental setup with hyperparameters or system-level training settings.