Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Frequency Matters: Explaining Biases of Face Recognition in the Frequency Domain

Marco Huber, Fadi Boutros, Naser Damer

TL;DR

This extensive results show that different frequencies are important to FR models depending on the ethnicity of the samples, and motivated by recent discoveries of the importance of frequency patterns in convolutional neural networks, bias in face recognition is explained using state-of-the-art frequency-based explanations.

Abstract

Face recognition (FR) models are vulnerable to performance variations across demographic groups. The causes for these performance differences are unclear due to the highly complex deep learning-based structure of face recognition models. Several works aimed at exploring possible roots of gender and ethnicity bias, identifying semantic reasons such as hairstyle, make-up, or facial hair as possible sources. Motivated by recent discoveries of the importance of frequency patterns in convolutional neural networks, we explain bias in face recognition using state-of-the-art frequency-based explanations. Our extensive results show that different frequencies are important to FR models depending on the ethnicity of the samples.

Frequency Matters: Explaining Biases of Face Recognition in the Frequency Domain

TL;DR

This extensive results show that different frequencies are important to FR models depending on the ethnicity of the samples, and motivated by recent discoveries of the importance of frequency patterns in convolutional neural networks, bias in face recognition is explained using state-of-the-art frequency-based explanations.

Abstract

Face recognition (FR) models are vulnerable to performance variations across demographic groups. The causes for these performance differences are unclear due to the highly complex deep learning-based structure of face recognition models. Several works aimed at exploring possible roots of gender and ethnicity bias, identifying semantic reasons such as hairstyle, make-up, or facial hair as possible sources. Motivated by recent discoveries of the importance of frequency patterns in convolutional neural networks, we explain bias in face recognition using state-of-the-art frequency-based explanations. Our extensive results show that different frequencies are important to FR models depending on the ethnicity of the samples.

Paper Structure

This paper contains 20 sections, 7 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Bias in Face Recognition
  4. Explainable Face Recognition
  5. Methodology
  6. Frequency-based Explanations
  7. Amplifying the Ethnicity Bias of FR Models
  8. Relative Frequency Ranking
  9. Impact of Bias on Frequency Importance
  10. Experimental Setup
  11. Training and Testing Dataset
  12. Face Recognition Models
  13. Bias Evaluation Metrics
  14. Frequency Explanations Parameters
  15. Results
  16. ...and 5 more sections

Figures (5)

  • Figure 1: Relative frequency ranking of the ethnicity for all models. Different pattern can be observed, such as low importance of high frequencies for African (red) or high importance of high frequencies for Asian (orange) in the most cases compared to the other ethnicities.
  • Figure 2: Mean frequency importance distribution based on ethnicity for model $M$
  • Figure 3: Mean frequency importance distribution based on Ethnicity for model $M_{\overline{Afr}}$ and $M_{\overline{Asi}}$. The striped bar indicate the distribution of the baseline $M$, brightness changes indicate differences between $M$ and the biased models.
  • Figure 4: Mean frequency importance distribution based on Ethnicity for model $M_{\overline{Cau}}$ and $M_{\overline{Ind}}$. The striped bar indicate the distribution of the baseline $M$, brightness changes indicate differences between $M$ and the biased models.
  • Figure 5: Mean frequency importance distribution based on ethnicity for the two state-of-the-art models.