Research map / Probabilistic and scientific modeling / Molecules and proteins
Obliviator Reveals the Cost of Nonlinear Guardedness in Concept Erasure
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 Obliviator, a post-hoc iterative kernel (HSIC/RKHS) concept erasure method that guards against nonlinear adversaries and traces the utility-erasure trade-off curve over iterations.
Problem
Concept erasure removes unwanted attributes (e.g., gender, race) from learned representations while preserving task utility, but existing methods remain vulnerable to nonlinear adversaries and their cost of erasure over the process is unstudied.
Approach
Formulates erasure functionally via HSIC in RKHS, minimizing dependence between representation and unwanted attribute while keeping task information visible through witness functions; solves the nested problem iteratively with an encoder step and a closed-form eigenvalue (RKHS disentanglement) step, yielding full utility-erasure trade-off curves.
Key results
- On Dial-Mention, Dial-Sentiment and Bias in Bios, the best baseline leaves a gap of 12%, 13% and 14% respectively between full erasure (chance) and the best-performing baseline in the BERT setting (per the paper's claim).
- Obliviator consistently achieves higher task (Y) accuracy at every level of attribute (S) accuracy across frozen/fine-tuned and supervised/unsupervised settings, per the paper.
- Erasure becomes more utility-preserving on better-disentangled representations: LLaMA and DeepSeek representations give higher trade-off curves than BERT on Bias in Bios.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Obliviator instead of existing erasure baselines | BERT representations · Bias in Bios, Dial-Mention, and Dial-Sentiment | target-task accuracy at a given sensitive-attribute accuracy: 12%, 13%, and 14% gap between full erasure and the best baseline on Dial-Mention, Dial-Sentiment, and Bias in Bios, resp | Helped |
| Use supervised erasure with target-task labels rather than unsupervised erasure | Obliviator · Dial-Mention and Bias in Bios | target-task accuracy during erasure | Helped |
| Use frozen GPT-2 representations instead of frozen BERT representations | Obliviator · Bias in Bios | utility-erasure trade-off | No clear effect |
| Use frozen LLaMa representations instead of frozen BERT representations | Obliviator · Bias in Bios | target-task utility at a given erasure level | Helped |
| Use frozen DeepSeek representations instead of frozen BERT representations | Obliviator · Bias in Bios, unsupervised erasure | utility-erasure trade-off | Helped |
| Increase the sampling skew of the unwanted attribute | Obliviator · Dial-Sentiment, DeepMoji, unsupervised erasure | utility-erasure trade-off | Hurt |
| Apply Obliviator erasure | BERT and DeepSeek representations · Dial-Sentiment | fairness metrics (Demographic Parity and Gap_rms) | Helped |
| Use supervised rather than unsupervised erasure | Obliviator · Dial-Sentiment fairness evaluation | utility at a given unfairness level | Helped |
Surprising findings
- Unsupervised erasure shows a more noticeable utility-erasure trade-off than supervised erasure, and fine-tuning PLMs does not eliminate nonlinear leakage for existing methods.
- Skewed sampling of the protected attribute within task classes worsens the trade-off because the HSIC estimate becomes biased.
Limitations
- Post-hoc HSIC estimates depend on how well the erasure dataset represents the true distribution; skewed samples worsen the trade-off.
- Requires choosing hyperparameters (tau weights, eigenvalue threshold) and encoder epochs; performance depends on these choices.
- Experiments are limited to a few text datasets and PLMs.
Experimental setup
Datasets: Bias in Bios Dial-Sentiment Dial-Mention
Benchmarks: Bias in Bios Dial-Sentiment Dial-Mention
Metrics: probe accuracy for protected attribute task accuracy demographic parity gap_rms
Baselines: INLP AdS kSAL FaRM KRaM
Builds on: SAL KCE / R-LACE kSAL KRaM FaRM
Closest papers by meaning
- Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them (ICLR 2025 · similarity 0.86)
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.86)
- Erasing Conceptual Knowledge from Language Models (NeurIPS 2025 · similarity 0.86)
- EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025 · similarity 0.86)
- When Are Concepts Erased From Diffusion Models? (NeurIPS 2025 · similarity 0.86)
- Consistency-Preserving Concept Erasure via Unsafe–Safe Pairing and Directional Fisher-weighted Adaptation (NeurIPS 2026 · similarity 0.85)
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.