Abstract: The widespread adoption of machine learning systems has raised critical concerns about fairness and bias, making mitigating harmful biases essential for AI development. In this paper, we investigate the relationship between debiasing and removing artifacts in neural networks for computer vision tasks. First, we introduce a set of novel XAI-based metrics that analyze saliency maps to assess shifts in a model's decision-making process. Then, we demonstrate that successful debiasing methods systematically redirect model focus away from protected attributes. Finally, we show that techniques originally developed for artifact removal can be effectively repurposed for improving fairness. These findings provide evidence for the existence of a bidirectional connection between ensuring fairness and removing artifacts corresponding to protected attributes.
1. Introduction
Deep learning models are notorious for finding "shortcuts." In computer vision, a model tasked with identifying a "smiling" person might accidentally learn that "wearing a necktie" is a strong predictor because of biases in the training data. In this scenario, the necktie acts as a proxy for a protected attribute (gender), leading to discriminatory outcomes.
While metrics like Equalized Odds measure if a model is biased, they often fail to explain how internal reasoning changes during debiasing. This paper uncovers a deep, bidirectional link between AI fairness and "artifact removal." By introducing new metrics to analyze saliency maps, we demonstrate that effective debiasing methods work by physically shifting a model's attention away from "protected" features (like a necktie) and toward relevant ones (like facial expressions).
2. Methodology
We approach the problem by treating protected attributes as "artifacts" that need to be removed from the model's decision-making process. We propose four novel XAI metrics to quantify how much a model relies on a specific Region of Interest (ROI):
- Rectangle Relevance Fraction (RRF): The percentage of total model "attention" falling within the ROI.
- Average Difference in Region (ADR): The decrease in saliency intensity in the ROI after debiasing.
- Decreased Intensity Fraction (DIF): The percentage of pixels in the ROI that saw a drop in importance.
- Rectangle Difference Distribution Testing (RDDT): A statistical t-test to confirm if the shift in attention is significant.
We compared three types of debiasing strategies: Threshold Optimizers (ThrOpt), Adversarial Fine-tuning (ZhangAL, SavaniAFT), and Concept-based Interventions (ClArC)—the latter being techniques originally designed for artifact removal.
3. Experimental Setup
The study utilizes the CelebA dataset with "Smiling" as the target attribute and "WearingNecktie" or "WearingHat" as protected attributes. We artificially controlled the correlation (using Yule’s $\phi$ coefficient) between the protected attributes and the target to simulate varying levels of dataset bias.
4. Results and Analysis
Our analysis reveals a significant "reasoning gap" in traditional methods. Simple thresholding (ThrOpt) improves fairness scores on paper but does not change the model's reasoning. Saliency maps show the model still focuses on the necktie; it merely changes how it interprets that information.
In contrast, methods like ClArC and ZhangAL showed a clear shift in saliency maps—moving the positive relevance from the necktie area to the eyes and mouth. ClArC variants were particularly aggressive, almost completely eliminating the model's reliance on the necktie region.
These results suggest a bidirectional connection: methods optimizing for fairness naturally redirect saliency away from protected attributes, and methods that redirect saliency naturally improve fairness. This implies that "bias" in vision models is often just a specific type of "artifact" or "Clever Hans" feature.
5. Conclusion and Future Work
This work provides evidence that fairness and artifact removal are two sides of the same coin. While effective, the current metrics rely on rectangular ROIs, which may not capture more complex protected attributes like skin tone or age-related textures. Future research should explore non-rectangular masks and address the accuracy trade-offs observed in the most effective debiasing methods. Expanding this framework to diverse domains like medical imaging or autonomous driving will further validate the link between fairness and artifact removal.
Summary updated