Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MST-KD: Multiple Specialized Teachers Knowledge Distillation for Fair Face Recognition

Eduarda Caldeira, Jaime S. Cardoso, Ana F. Sequeira, Pedro C. Neto

TL;DR

This paper tackles fairness in face recognition by addressing cross-ethnicity performance gaps. It introduces MST-KD, a multi-teacher knowledge distillation framework where four ethnicity-specific teachers train distinct embeddings that are projected into a common multi-teacher space and distilled to a single student using either a-KD or EAF-KD losses, with ElasticArcFace guiding classification. Empirical results show that distilling from specialized teachers improves overall accuracy and reduces bias (as measured by STD and SER) compared to baselines and generalist setups, with the simplest adaptor (SL) often yielding the best trade-offs. The approach is privacy-friendly (no ethnicity identifiers needed by the student during distillation) and scalable, though current experiments are limited by dataset size; future work includes larger and synthetic datasets, larger backbones, and more teachers to further enhance fairness and performance.

Abstract

As in school, one teacher to cover all subjects is insufficient to distill equally robust information to a student. Hence, each subject is taught by a highly specialised teacher. Following a similar philosophy, we propose a multiple specialized teacher framework to distill knowledge to a student network. In our approach, directed at face recognition use cases, we train four teachers on one specific ethnicity, leading to four highly specialized and biased teachers. Our strategy learns a project of these four teachers into a common space and distill that information to a student network. Our results highlighted increased performance and reduced bias for all our experiments. In addition, we further show that having biased/specialized teachers is crucial by showing that our approach achieves better results than when knowledge is distilled from four teachers trained on balanced datasets. Our approach represents a step forward to the understanding of the importance of ethnicity-specific features.

MST-KD: Multiple Specialized Teachers Knowledge Distillation for Fair Face Recognition

TL;DR

This paper tackles fairness in face recognition by addressing cross-ethnicity performance gaps. It introduces MST-KD, a multi-teacher knowledge distillation framework where four ethnicity-specific teachers train distinct embeddings that are projected into a common multi-teacher space and distilled to a single student using either a-KD or EAF-KD losses, with ElasticArcFace guiding classification. Empirical results show that distilling from specialized teachers improves overall accuracy and reduces bias (as measured by STD and SER) compared to baselines and generalist setups, with the simplest adaptor (SL) often yielding the best trade-offs. The approach is privacy-friendly (no ethnicity identifiers needed by the student during distillation) and scalable, though current experiments are limited by dataset size; future work includes larger and synthetic datasets, larger backbones, and more teachers to further enhance fairness and performance.

Abstract

As in school, one teacher to cover all subjects is insufficient to distill equally robust information to a student. Hence, each subject is taught by a highly specialised teacher. Following a similar philosophy, we propose a multiple specialized teacher framework to distill knowledge to a student network. In our approach, directed at face recognition use cases, we train four teachers on one specific ethnicity, leading to four highly specialized and biased teachers. Our strategy learns a project of these four teachers into a common space and distill that information to a student network. Our results highlighted increased performance and reduced bias for all our experiments. In addition, we further show that having biased/specialized teachers is crucial by showing that our approach achieves better results than when knowledge is distilled from four teachers trained on balanced datasets. Our approach represents a step forward to the understanding of the importance of ethnicity-specific features.
Paper Structure (20 sections, 7 equations, 3 figures, 6 tables)

This paper contains 20 sections, 7 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Bias
  4. Knowledge Distillation
  5. State of the Art
  6. Bias Mitigation
  7. Knowledge Distillation
  8. Knowledge Distillation for Bias Mitigation
  9. Methodology
  10. Data
  11. Teachers
  12. Space Adaptors
  13. Students
  14. Baseline
  15. Experimental Setup
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Different data splits performed before teacher training for our (top) and the baseline (bottom) multi-teacher approaches.
  • Figure 2: Training process of the proposed adaptor networks. Note that the previously trained teacher models are frozen during this process.
  • Figure 3: Training process of the proposed student networks using EAF-KD. For a-KD students, the classification loss term is disregarded; thus, no ground truth labels are required during training. Note that the previously trained teachers and adaptor are frozen during this process.