Abstract
Although the volume of literature and public attention on machine learning fairness has been growing significantly, in practice some tasks as basic as measuring fairness, which is the first step in studying and promoting fairness, can be challenging. This is because sensitive attributes are often unavailable due to privacy regulations. The straightforward solution is to use auxiliary models to predict the missing sensitive attributes. However, our theoretical analyses show that the estimation error of the directly measured fairness metrics is proportional to the error rates of auxiliary models' predictions. Existing works that attempt to reduce the estimation error often require strong assumptions, e.g. access to the ground-truth sensitive attributes or some form of conditional independence. In this paper, we drop those assumptions and propose a framework that uses only off-the-shelf auxiliary models. The main challenge is how to reduce the negative impact of imperfectly predicted sensitive attributes on the fairness metrics without knowing the ground-truth sensitive attributes. Inspired by the noisy label learning literature, we first derive a closed-form relationship between the directly measured fairness metrics and their corresponding ground-truth metrics. And then we estimate some key statistics (most importantly transition matrix in the noisy label literature), which we use, together with the derived relationship, to calibrate the fairness metrics. In addition, we theoretically prove the upper bound of the estimation error in our calibrated metrics and show our method can substantially decrease the estimation error especially when auxiliary models are inaccurate or the target model is highly biased. Experiments on COMPAS and CelebA validate our theoretical analyses and show our method can measure fairness significantly more accurately than baselines under favorable circumstances.
Introduction
To ensure algorithmic fairness, researchers typically calculate metrics such as Demographic Parity (DP) or Equalized Odds (EOd). These calculations require access to sensitive attributes (e.g., race, gender), which are frequently unavailable due to privacy regulations like GDPR. A common industry workaround is to use "proxy models" to predict these missing attributes. However, as shown in Figure 1, naive use of these proxies can be dangerously misleading. The fairness disparity measured using proxies is often significantly lower than the true disparity, creating a "false sense of fairness" while the underlying model remains biased.
Methodology: Calibration via Weak Proxies
Instead of attempting to build perfect proxies—which would itself pose a privacy risk—this work focuses on calibration. The proposed framework treats proxy outputs as "noisy labels" of the true sensitive attribute.
The Calibration Pipeline
The framework (illustrated in Figure 2) utilizes a closed-form matrix relationship to "undo" the noise introduced by proxies. The core challenge is estimating the "transition matrix" (the probability of a proxy being incorrect) without access to ground-truth data.
Higher-Order Consensus (HOC)
To achieve calibration without ground truth, the authors adapt techniques from noisy label learning. By employing at least three independent proxies, the algorithm analyzes patterns of agreement and disagreement. This "Higher-Order Consensus" allows for the mathematical estimation of the noise distribution (the transition matrix) without needing the truth.
Key Insights and Results
The research yields a counter-intuitive finding: weak proxies are often preferable. Highly accurate proxies effectively reconstruct sensitive data, potentially violating privacy. Conversely, weak proxies (e.g., ~60% accuracy) provide a "layer of protection" for user privacy while still allowing for near-perfect aggregate fairness measurements when calibrated.
Experimental Performance
- Accuracy Recovery: On the COMPAS dataset (recidivism prediction), the calibration algorithm was 39.6% to 88.2% more accurate at measuring fairness than directly using proxies.
- Robustness to Noise: On the CelebA dataset (facial attributes), even when proxies were intentionally weakened to ~58% accuracy to protect privacy, the calibrated error remained low, whereas baseline errors increased tenfold.
- Bias Mitigation: When used as a constraint during training, the calibrated metrics allowed models to achieve fairness levels nearly identical to those trained with full access to sensitive attributes.
Limitations and Future Work
The framework's success relies on the relative independence of proxy errors. If all proxies rely on the same biased features (e.g., identical census data), calibration may fail. Additionally, while the method works well for binary attributes, scaling to multi-class sensitive attributes increases the complexity of the transition matrix estimation. Future work will explore refining the selection between "Global" and "Local" noise estimation heuristics.
Summary updated