Research map / Learning theory and algorithms / Fairness
An Information-Flow Perspective on Algorithmic Fairness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Links secure information flow (noninterference, conditional vulnerability) to algorithmic fairness, proposing a quantitative "fairness spread" metric analyzed with off-the-shelf tools (KeY, Joana, counterSharp) on programs and causal models.
Problem
Algorithmic fairness is usually evaluated by statistical metrics treating the algorithm as a black box, missing disparities in the decision-making process; the paper asks how information-flow analysis of program code can formalize and measure fairness.
Approach
Treat protected attributes as high-security inputs so that noninterference equals demographic parity/fairness; define a quantitative fairness spread via conditional vulnerability under a uniform group distribution, shown equal to a counterfactual-fairness-related bound, and analyzable by model counting on program code.
Key results
- Fairness spread for the causal credit example: c3 has 0.23 vs c2 0.27 under the causal model, reversing the non-causal ranking where c2 is fairer.
- Car insurance case study: the engine-only classifier has a fairness spread of 0.36 while the engine+gender classifier is counterfactually fair (spread 0).
- Conditional vulnerability equals (|G|·V − 1) = fairness spread under uniform G, and the paper proves Pr[deviating counterfactual] ≤ fairness spread, with equality for two groups.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a non-uniform credit-score distribution, increasing the share of scores 6–7 | c3 · Credit decision examples | Fairness Spread: 0.2 -> 0.3 | Hurt |
| Use c2 instead of c1 | Credit decision examples · Uniform credit-score distribution | Fairness Spread: 1.0 -> 0.0 | Helped |
| Use c3 instead of c1 | Credit decision examples · Uniform credit-score distribution | Fairness Spread: 1.0 -> 0.2 | Helped |
| Use c2 instead of c3 | Credit decisions under the causal model · Credit decisions under causal model | Fairness Spread: 0.23 -> 0.27 | Hurt |
| Use c3 instead of c2 | Credit decisions under the causal model · Credit decisions under causal model | Fairness Spread: 0.27 -> 0.23 | Helped |
| Hold zipCode fixed across counterfactual evaluations | c3 · Path-specific credit decisions under causal model | Fairness Spread: 0.23 -> 0.19 | Helped |
| Hold zipCode fixed across counterfactual evaluations | c2 · Path-specific credit decisions under causal model | Fairness Spread: 0.27 -> 0 | Helped |
| Use both engine and gender instead of engine only | Car insurance classifier · Car insurance decisions under causal model | Counterfactual fairness | Helped |
| Use engine only rather than both engine and gender | Car insurance classifier · Car insurance decisions under causal model | Fairness Spread: 0 -> 0.36 | Hurt |
| Analyze German wage-tax computation with information-flow analysis | German wage-tax program · German wage-tax computation (2015–2023) | Disparate treatment | No clear effect |
Surprising findings
- Under a causal model with zipCode depending on group, the ranking of c2 and c3 by fairness flips relative to the independent-variable analysis.
- A corrigendum reports that the restricted information flow implies conditional demographic parity lemma is false, invalidating that part of the paper.
Limitations
- Assumes independence of protected group and unprotected attributes in the core results; dependence handled only via causal models.
- Restricted information flow lemma on conditional demographic parity is invalid (corrigendum).
- Fairness spread is defined for binary decisions and deterministic programs; analysis uses finite, quantized inputs.
Experimental setup
Metrics: fairness spread conditional vulnerability demographic parity counterfactual fairness
Baselines: conditional vulnerability (naive fairness metric) demographic parity
Builds on: noninterference quantitative information flow counterfactual fairness demographic parity path-specific counterfactual fairness
Closest papers by meaning
- [Re] Exacerbating Algorithmic Bias through Fairness Attacks (NeurIPS 2022 · similarity 0.84)
- Characterizing the risk of fairwashing (NeurIPS 2021 · similarity 0.84)
- Fairness under Competition (NeurIPS 2025 · similarity 0.83)
- Procedural Fairness Through Decoupling Objectionable Data Generating Components (ICLR 2024 · similarity 0.83)
- Searching for Unfairness in Algorithms’ Outputs: Novel Tests and Insights (AAAI 2025 · similarity 0.83)
- Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.