Gone With the Bits: Revealing Racial Bias in Low-Rate Neural Compression for Facial Images

2025-05-05

AI Summary

This work identifies, for the first time, systematic reconstruction bias in neural image compression models toward facial images of different racial groups at ultra-low bitrates (<0.1 bpp). We propose the first scalable evaluation framework for racial bias in image compression, covering nine state-of-the-art models and their variants. We demonstrate that conventional distortion metrics (e.g., PSNR, MS-SSIM) fail to capture this bias. Through ablation studies—including facial phenotypic degradation analysis, cross-model benchmarking, and balanced-data training—we confirm statistically significant bias across all tested models; balanced training only partially mitigates it. We further quantify the respective contributions of bias originating from the compression backbone versus downstream classification modules. Finally, we reveal an inherent trade-off between perceptual fidelity and fairness, establishing both theoretical foundations and empirical evidence for fairness-aware visual compression modeling.

Technology Category

• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Neural compression methods are gaining popularity due to their superior rate-distortion performance over traditional methods, even at extremely low bitrates below 0.1 bpp. As deep learning architectures, these models are prone to bias during the training process, potentially leading to unfair outcomes for individuals in different groups. In this paper, we present a general, structured, scalable framework for evaluating bias in neural image compression models. Using this framework, we investigate racial bias in neural compression algorithms by analyzing nine popular models and their variants. Through this investigation, we first demonstrate that traditional distortion metrics are ineffective in capturing bias in neural compression models. Next, we highlight that racial bias is present in all neural compression models and can be captured by examining facial phenotype degradation in image reconstructions. We then examine the relationship between bias and realism in the decoded images and demonstrate a trade-off across models. Finally, we show that utilizing a racially balanced training set can reduce bias but is not a sufficient bias mitigation strategy. We additionally show the bias can be attributed to compression model bias and classification model bias. We believe that this work is a first step towards evaluating and eliminating bias in neural image compression models.

Problem

• Evaluating racial bias in neural image compression models
• Identifying ineffective traditional distortion metrics for bias detection
• Exploring bias-realism trade-off in decoded facial images

Innovation

• Framework for evaluating bias in neural compression
• Analyzes racial bias via facial phenotype degradation
• Uses racially balanced training set to reduce bias