Research map / Label-efficient and robust learning / Recognition and long-tail
FAIR-FER: A Latent Alignment Approach for Mitigating Bias in Facial Expression Recognition (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FAIR-FER, a shared-weight VAE with an adversarial discriminator aligning latent spaces across protected attributes, followed by an MBConv classifier, to reduce demographic bias in facial expression recognition on RAF-DB.
Problem
Human-annotated facial expression recognition datasets propagate annotator and demographic biases (gender, race, age), reducing equal-odds fairness of FER models.
Approach
Encode images of each protected attribute into a shared-weight VAE latent space, use an adversarial discriminator to align the latent spaces across attributes so they carry no attribute information, then classify the aligned latent with an MBConv-based module.
Key results
- Mean expression accuracy on RAF-DB of 76.1% vs 74.8% for Xu et al. 2020.
- Bias mitigation for race improved to 94.2 from 91.6 and for age to 84.8 from 82.1; gender slightly lower at 99.51 vs 99.97.
- Fear class accuracy improved from 53.8% to 60.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FAIR-FER instead of Xu et al. | FER classification · RAF-DB test, Mean | mean accuracy (%): 74.8 -> 76.1 | Helped |
| Use FAIR-FER instead of Xu et al. | FER classification · RAF-DB test, expression-wise accuracy | accuracy (%): 81.0 -> 83.2; 54.1 -> 57.7; 53.8 -> 60.2; 93.3 -> 92.0; 82.1 -> 81.0; 77.7 -> 76.0; 81.8 -> 82.9 | Mixed |
| Use FAIR-FER instead of Xu et al. | FER bias mitigation · RAF-DB test, Gender | mitigation of bias: 99.97 -> 99.51 | Hurt |
| Use FAIR-FER instead of Xu et al. | FER bias mitigation · RAF-DB test, Race | mitigation of bias: 91.6 -> 94.2 | Helped |
| Use FAIR-FER instead of Xu et al. | FER bias mitigation · RAF-DB test, Age | mitigation of bias: 82.1 -> 84.8 | Helped |
Surprising findings
- Gender fairness score drops slightly (99.97 to 99.51) relative to the baseline, despite gains on race and age.
Limitations
- Evaluated only on RAF-DB with a single baseline comparison.
- Happy class accuracy decreased relative to baseline (92.0 vs 93.3).
Experimental setup
Datasets: RAF-DB
Benchmarks: RAF-DB
Metrics: per-class accuracy mean per-class per-attribute accuracy equal odds fairness ratio (min over attributes)
Baselines: Xu et al. 2020
Builds on: variational autoencoder EfficientNet MBConv Xu et al. 2020 fairness FER
Closest papers by meaning
- Open-Set Facial Expression Recognition (AAAI 2024 · similarity 0.86)
- Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis (AAAI 2026 · similarity 0.86)
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · similarity 0.85)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.84)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.84)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.