A Logical Framework for Verification of AI Fairness

Priya T V, Shrisha Rao

Primary Area: societal considerations including fairness, safety, privacy

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Artificial Intelligence, bias, discrimination, fairness, spider diagram, first-order predicate logic

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: A framework to define and verify fairness of an AI model using the spider diagrams.

Abstract: With the widespread use of AI in socially important decision-making processes, it becomes crucial to ensure that AI-generated decisions do not reflect discrimination towards certain groups or populations. To address this challenge, our research introduces a theoretical framework based on the spider diagram---a reasoning system rooted in first-order predicate logic, and an extended version of the Euler and Venn diagrams---to define and verify the fairness of AI algorithms in decision-making. This framework compares the sets representing the actual outcome of the algorithm and the expected outcome to identify bias in the model. The expected outcome of the model is calculated by considering the similarity score between the individual instances in the dataset. If the set of actual outcomes is a subset of the set of expected outcomes and all constant spiders in the former set have a corresponding foot in the expected outcome set, then the algorithm is free from bias. We further evaluate the performance of the AI model using the spider diagram that replaces the conventional confusion matrix in the literature. The framework also permits us to define a degree of bias and evaluate the same for specific AI models. Experimental results indicate that this framework surpasses traditional approaches in efficiency, with improvements in processing time and a reduced number of function calls.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 839