An Information-Flow Perspective on Algorithmic Fairness

Authors: Samuel Teuber, Bernhard Beckert

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

Reproducibility Variable Result LLM Response
Research Type Experimental This work presents insights gained by investigating the relationship between algorithmic fairness and the concept of secure information flow. [...] We demonstrate that off-the-shelf tools for information-flow properties can be used in order to formally analyze a program s algorithmic fairness properties, including the new notion of fairness spread as well as established notions such as demographic parity. [...] Throughout this paper, we will demonstrate our ideas using the three exemplary programs shown in Figure 1: All programs make a loan allocation decision based on a protected group attribute (group) and an unprotected variable containing the credit score (score). [...] We applied our approach to two case-studies, showcasing both the qualitative (Section 6.1) as well as the quantitative (Section 6.2) approach. [...] All analyses are summarized in Table 1 and were computed using the counter Sharp tool which allows the quantitative analysis of C-programs.
Researcher Affiliation Academia Karlsruhe Institute of Technology, Am Fasanengarten 5, 76131 Karlsruhe {teuber,beckert}@kit.edu
Pseudocode Yes Figure 1: Three exemplary credit decision algorithms: 1 def c1(group, score): 2 return (group != 0) (a) Group Decision; 1 def c2(group, score): 2 return (score>=8) (b) Score Decision; 1 def c3(group, score): 2 if (group >= 6): 3 return (score >= 8) 4 else: 5 return (score >= 6) (c) Mixed Decision
Open Source Code Yes Artifact available at (Teuber and Beckert 2023c). [...] Teuber, S.; and Beckert, B. 2023c. samysweb/AAAI24Fairness: AAAI24 Artefact. 10.5281/zenodo.10385360.
Open Datasets No The paper defines illustrative examples and distributions for its analyses (e.g., 'a setting with 10 groups (e.g., age ranges) and a credit score of U between 1 and 10', 'Ud is the discrete uniform distribution'). However, it does not use or provide access information for a named, publicly available dataset in the traditional sense of a common benchmark or repository.
Dataset Splits No The paper does not provide explicit training/test/validation dataset splits. Its analyses are based on formal methods and properties of programs with defined inputs/distributions, not on splitting a large dataset into subsets for training, validation, and testing of a model.
Hardware Specification No The paper states that 'Each analysis took less than 3 seconds of computation time' in Table 1 but does not specify any hardware details like CPU or GPU models, memory, or specific computing environments used for the analyses.
Software Dependencies No The paper mentions several tools used: 'the interactive theorem prover Ke Y (Ahrendt et al. 2005, 2016)', 'the automated information-flow analysis tool Joana (Graf, Hecker, and Mohr 2013; Snelting et al. 2014)', and 'the tool counter Sharp (Teuber and Weigl 2021)'. It also refers to a 'Java implementation (Federal Ministry of Finance 2022a)'. While these tools are named and cited, no specific version numbers are provided for any of them.
Experiment Setup Yes To this end, we analyzed all three programs for a setting with 10 groups (e.g., age ranges) and a credit score of U between 1 and 10. Initially, we assumed a uniform distribution of credit scores. In addition, we analyzed c3 (Figure 1 (b)) using a wrapper program that adjusted the probability of a credit score between 6 to 7 to be 30% (instead of 20% in the uniform case). [...] We assume B1 Ud (0, 9), B2 Ud (0, 9), B3 Ud ( 1, 5), B4 Ud ( 3, 3) where Ud is the discrete uniform distribution.