Research map / Label-efficient and robust learning / Recognition and long-tail
MABR: Multilayer Adversarial Bias Removal Without Prior Bias Knowledge
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 MABR, an adversarial debiasing method that dynamically detects biased samples with per-layer bias detectors and misclassified examples, then removes them via reverse-gradient domain discriminators, without protected-attribute labels.
Problem
Social biases in NLP classifiers are typically mitigated only with prior knowledge of the bias type and protected-attribute (demographic) labels, which are often unavailable due to privacy and unknown bias types.
Approach
Attach auxiliary bias detectors to every encoder layer that predict whether the main model is correct without seeing the task label; samples flagged by detectors (plus misclassified training samples) are treated as biased and an adversarial reverse-gradient discriminator at each layer makes representations invariant to them, with detectors and discriminators trained jointly so labels are refined d
Key results
- On sentiment with BERT, MABR reduces TPR_RMS from 0.243 to 0.137 and independence from 0.039 to 0.003 with accuracy 0.766 vs 0.771 for finetuned.
- On occupation classification with BERT, MABR lowers TPR_RMS from 0.135 to 0.101 and independence from 0.149 to 0.099 with accuracy 0.857 vs 0.869.
- Ablation removing multi-layer bias detection increases TPR_RMS on sentiment BERT from 0.137 to 0.145 (per the main table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply MABR instead of standard fine-tuning | BERT · Sentiment analysis | TPR_RMS: 0.243 -> 0.137 | Helped |
| apply MABR instead of standard fine-tuning | BERT · Sentiment analysis | Indep: 0.039 -> 0.003 | Helped |
| apply MABR instead of standard fine-tuning | RoBERTa · Sentiment analysis | TPR_RMS: 0.261 -> 0.126 | Helped |
| apply MABR instead of standard fine-tuning | RoBERTa · Sentiment analysis | Indep: 0.035 -> 0.006 | Helped |
| apply MABR instead of standard fine-tuning | BERT · Occupation classification | TPR_RMS: 0.135 -> 0.101 | Helped |
| apply MABR instead of standard fine-tuning | BERT · Occupation classification | Indep: 0.149 -> 0.099 | Helped |
| apply MABR instead of standard fine-tuning | RoBERTa · Occupation classification | TPR_RMS: 0.132 -> 0.109 | Helped |
| apply MABR instead of standard fine-tuning | RoBERTa · Occupation classification | Indep: 0.144 -> 0.100 | Helped |
| add multi-layer bias detection (ablation) | MABR with BERT · Sentiment analysis | TPR_RMS: 0.145 -> 0.137 | Helped |
| add hard-example detection (ablation) | MABR with BERT · Sentiment analysis | TPR_RMS: 0.139 -> 0.137 | Helped |
| add multi-layer bias detection (ablation) | MABR with RoBERTa · Sentiment analysis | TPR_RMS: 0.162 -> 0.126 | Helped |
| add hard-example detection (ablation) | MABR with RoBERTa · Sentiment analysis | TPR_RMS: 0.142 -> 0.126 | Helped |
Surprising findings
- Sometimes outperforms methods that use demographic labels, despite never seeing protected attributes.
- Bias detectors remain highly accurate (above 0.76 for BERT) even after adversarial training, so dynamically refined labels are needed.
Limitations
- Relies on threshold tau and multiple hyperparameters; threshold chosen so ~30% of samples fall outside it.
- Evaluated only on two binary-attribute English tasks with BERT-family encoders.
- Model selection uses DTO criterion and accuracy threshold rather than demographic validation data, which may affect reported fairness.
Experimental setup
Datasets: 100,000 tweets dialect/sentiment dataset (AAE vs MUSE) 400,000 online biographies occupation dataset (Bias in Bios)
Benchmarks: sentiment analysis tweets dataset (race) occupation classification biographies dataset (gender)
Metrics: accuracy TPR RMS gap independence (KL) sufficiency (KL)
Baselines: Finetuned INLP R-LACE JTT BLIND JTT-Disc + AdvTrain Bias-Disc + Upweight
Builds on: INLP R-LACE JTT BLIND Reverse Gradient Layer
Closest papers by meaning
- BiasAdv: Bias-Adversarial Augmentation for Model Debiasing (CVPR 2023 · similarity 0.90)
- Interpretable Debiasing of Vision-Language Models for Social Fairness (CVPR 2026 · similarity 0.87)
- The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models (ICLR 2024 · similarity 0.87)
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.85)
- SANER: Annotation-free Societal Attribute Neutralizer for Debiasing CLIP (ICLR 2025 · similarity 0.85)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · 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.