FAIR-SIGHT: Fairness Assurance in Image Recognition via Simultaneous Conformal Thresholding and Dynamic Output Repair

TL;DR

FAIR-SIGHT tackles fairness in vision by delivering a post-hoc, model-agnostic solution with formal guarantees. It defines a fairness-aware non-conformity score $S(I) = d(h(I), y_{ref}(I)) + \lambda \Delta(I,A)$ and uses conformal prediction to calibrate a threshold $Q_\alpha$, ensuring at most an $\alpha$-fraction of outputs are unfair. When a sample exceeds $Q_\alpha$, targeted repairs such as logit shifts for classification and confidence recalibration for detection restore fairness without retraining, with an adaptive threshold update to handle distribution shifts. Empirically, the method reduces group disparities while preserving high accuracy across CelebA, UTKFace, and COCO-based datasets, demonstrating scalable, black-box fairness in high-dimensional vision tasks.

Abstract

We introduce FAIR-SIGHT, an innovative post-hoc framework designed to ensure fairness in computer vision systems by combining conformal prediction with a dynamic output repair mechanism. Our approach calculates a fairness-aware non-conformity score that simultaneously assesses prediction errors and fairness violations. Using conformal prediction, we establish an adaptive threshold that provides rigorous finite-sample, distribution-free guarantees. When the non-conformity score for a new image exceeds the calibrated threshold, FAIR-SIGHT implements targeted corrective adjustments, such as logit shifts for classification and confidence recalibration for detection, to reduce both group and individual fairness disparities, all without the need for retraining or having access to internal model parameters. Comprehensive theoretical analysis validates our method's error control and convergence properties. At the same time, extensive empirical evaluations on benchmark datasets show that FAIR-SIGHT significantly reduces fairness disparities while preserving high predictive performance.

Figures (4)

• Figure 1: High-level overview of proposed workflow. We start with input images (possibly containing a protected attribute), feed them into a black-box computer vision (CV) model, then apply FAIR-SIGHT module as a post-hoc fairness repair. Its output is a fair set of predictions, mitigating bias while preserving accuracy.
• Figure 2: FAIR-SIGHT Workflow. The Offline Calibration (left) takes a calibration dataset $(I_i, A_i)$, processes each sample to compute non-conformity scores ($S(I_i)$ or region-based $S_R(I_i,R_j)$), and derives the conformal fairness threshold(s) $\{Q_\alpha,\,Q_\alpha(R_j)\}$. These thresholds are stored for later use. The Online Inference (right) processes each new image $I_{\text{new}}$ through the trained computer vision (CV) model, computes $S(I_{\text{new}})$, and checks it against the stored thresholds. If $S(I_{\text{new}})$ (or any $S_R(I_{\text{new}},R_j)$) exceeds the threshold, we apply a post hoc correction (e.g., adjusting class logits or bounding-box confidences) and optionally update the threshold through the feedback loop. Otherwise, the raw model output is used as-is.
• Figure 3: Ablation study on FAIR-SIGHT hyperparameters (ResNet50, CelebA). Each panel plots fairness metrics (DPD/EOD) against a different parameter: $\lambda$ (error vs. fairness penalty), $\gamma$ (threshold update), $\eta$ (scaling factor), and $\kappa$ (logit shift aggressiveness). Middle-range values minimize disparities without harming accuracy.
• Figure 4: Qualitative Results.Top: t-SNE embeddings on UTKFace classification illustrate that baseline features cluster by protected attribute, whereas FAIR-SIGHT produces more intermingled clusters, indicating reduced bias in the feature space. Bottom: On the COCO-based detection dataset using a MambaVision-L2-1K model, a conformal calibration threshold of 0.73 is computed from the validation set. In the baseline output (left), several bounding boxes for persons in the protected group (race = Black) have confidence scores below 0.73 (indicated by red boxes), signaling under-detection for the protected group. After applying FAIR-SIGHT’s post-hoc repair mechanism, the scores of those protected group boxes are boosted so that detections meet the threshold, resulting in more balanced and fair outputs (right).