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

TL;DR

This work tackles racial bias in low-rate neural facial image compression by proposing a general, scalable framework to quantify bias beyond traditional distortion metrics. It defines a phenotype-degradation bias using a classifier-based metric and benchmarks nine neural compression models across multiple datasets, revealing bias that traditional metrics miss. The analysis demonstrates a bias-realism trade-off and shows that racially balanced training data can reduce but not eliminate bias, attributing discrimination to both compression- and classifier-induced effects. This study lays the groundwork for bias-aware design and mitigation of neural compression systems for facial images.

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.

Figures (42)

• Figure 2: Traditional rate-distortion metrics (PSNR, SSIM, and LPIPS) for the Joint model trained on the CelebA dataset, shown for each race and the overall dataset. The rate-distortion curves are nearly identical across all races for PSNR and SSIM, which contrasts with the findings from the qualitative analysis. While the LPIPS curve for the African group is slightly higher than for other races, it fails to fully reflect the disparities observed in the qualitative analysis.
• Figure 3: (a) Bias for Skin Type across different races for Joint reconstructions trained on the CelebA dataset. (b) As the bitrate is lowered, bias increases for Skin Type, Eye Type, and Hair Type, while remaining relatively level for other phenotypes.
• Figure 4: Bias in Skin Type and Eye Type across all neural compression models.
• Figure 5: At high bitrates ($>0.1$ bpp), bias and realism are correlated across all the models. At low bitrates ($<0.1$ bpp), the trend is more sporadic.
• Figure 6: (a) Using a racially balanced dataset (FaceARG) helps reduce the bias until extremely low bitrates less than 0.1 bpp. However, the general trend of bias increasing with decreasing bitrate is consistent across 2 datasets. (b) Using FairFace does not reduce bias and in cases increases bias.

Theorems & Definitions (1)

• Example 3.1