PaperlayerView exact revision on arXiv

Source-linked AI summary

Easy to Learn, Yet Hard to Forget: Towards Robust Unlearning Under Bias

JuneHyoung Kwon, MiHyeon Kim, Eunju Lee, Yoonji Lee, Seunghoon Lee, YoungBin Kim

arXiv:2602.21773v1cs.LGcs.CV

TL;DR

Biased models entangle causal and shortcut features, causing standard unlearning to remove shortcut reliance rather than the intended class information. CUPID partitions samples by sharpness, separates causal and bias pathways, and routes targeted updates; experiments on three biased datasets report superior, balanced forgetting. The method addresses this setting without requiring retain-set access, but its formulation assumes a biased training setup with spurious attributes correlated with class labels.

  • Problem

    Standard unlearning struggles when target-class information is entangled with spurious shortcuts, and existing methods may require infeasible retain-set access.

  • Method

    CUPID uses sharpness-aware partitioning, pathway identification, and targeted causal and bias gradients to selectively update model parameters.

  • Results

    CUPID achieves superior forgetting across Waterbirds, BAR, and Biased NICO++, including 3.75% FA on BAR versus 30.26% for the next-best method.

  • Takeaways & Limitations

    CUPID mitigates shortcut unlearning by selectively forgetting causal information while maintaining balanced performance across bias-aligned and bias-conflicting samples.

  • Takeaways & Limitations

    The formulation assumes biased data in which each sample has a spurious attribute highly correlated with its class label.

Abstract

from arXiv · show

Machine unlearning, which enables a model to forget specific data, is crucial for ensuring data privacy and model reliability. However, its effectiveness can be severely undermined in real-world scenarios where models learn unintended biases from spurious correlations within the data. This paper investigates the unique challenges of unlearning from such biased models. We identify a novel phenomenon we term ``shortcut unlearning," where models exhibit an ``easy to learn, yet hard to forget" tendency. Specifically, models struggle to forget easily-learned, bias-aligned samples; instead of forgetting the class attribute, they unlearn the bias attribute, which can paradoxically improve accuracy on the class intended to be forgotten. To address this, we propose CUPID, a new unlearning framework inspired by the observation that samples with different biases exhibit distinct loss landscape sharpness. Our method first partitions the forget set into causal- and bias-approximated subsets based on sample sharpness, then disentangles model parameters into causal and bias pathways, and finally performs a targeted update by routing refined causal and bias gradients to their respective pathways. Extensive experiments on biased datasets including Waterbirds, BAR, and Biased NICO++ demonstrate that our method achieves state-of-the-art forgetting performance and effectively mitigates the shortcut unlearning problem.

Introduction

The paper shows that biased models make unlearning unreliable because causal and shortcut features become entangled. It introduces CUPID, which separates these pathways and targets updates to improve class forgetting under bias.

  • Biased models exploit spurious correlations as shortcuts, entangling class-defining features with bias features.
  • Shortcut unlearning occurs when models erase shortcut reliance instead of the causal information defining the target class.
  • The paper identifies an easy-to-learn, hard-to-forget asymmetry for bias-aligned samples and paradoxical accuracy improvements on bias-conflicting samples.
  • CUPID uses loss-landscape sharpness to partition forget samples, disentangle causal and bias pathways, and route selective updates.
  • Experiments on Waterbirds, BAR, and Biased NICO++ report lower forget accuracy and balanced performance across bias-aligned and bias-conflicting samples.
  • The contributions formalize shortcut unlearning, introduce CUPID, and empirically validate selective forgetting of causal information on biased datasets.

Related Work

Machine unlearning research spans exact retraining and approximate techniques that reduce computational cost. Existing methods remain challenged by entangled target information and, often, unavailable retain-set access.

  • Exact unlearning retrains a model from scratch as a theoretical gold standard, but its computational cost is often impractical.
  • Approximate methods degrade forget-set performance by maximizing loss, injecting adversarial noise, or relabeling data.
  • Existing methods often assume cleanly separable target information, an assumption that fails when target classes correlate with task-irrelevant features.
  • Many existing methods also require retain-set access, which can be infeasible because of privacy or storage constraints.

Preliminary

The paper formulates class-wise forgetting in biased data, where spurious attributes correlate strongly with class labels. It distinguishes forget-set samples by whether shortcut attributes align with or conflict with the target class.

  • Class-wise forgetting removes the influence of a target class from an original model while producing an unlearned model.
  • Under data bias, each sample has a spurious attribute highly correlated with its class label, encouraging the model to learn a shortcut.
  • Bias-aligned samples have shortcuts that agree with the target class, whereas bias-conflicting samples have shortcuts that fail to predict it.
  • Figure 1 summarizes rapid learning of bias-aligned samples, their slower forgetting, shortcut removal by linear probing, and sharpness differences between sample types.

Analysis: Unlearning on Biased Models

Standard unlearning on biased models learns bias-aligned samples quickly but forgets them slowly, while bias-conflicting samples are forgotten more readily. The resulting shortcut removal can improve predictions for samples in the class meant to be forgotten.

  • During training, models first learn easier patterns from bias-aligned samples before fitting harder bias-conflicting samples.
  • Bias-conflicting samples attain substantially larger forgetting-loss magnitudes than bias-aligned samples under NegGrad, indicating easier forgetting.
  • The forgetting loss for bias-conflicting samples sharply decreases, producing a paradoxical debiasing effect in their predictions.
  • These dynamics indicate that standard unlearning removes dominant shortcut representations rather than the intended causal features.
  • Accuracy on the bias-conflicting forget subset increases after unlearning, while a linear probe shows reduced bias-label accuracy in frozen representations.

Method

CUPID uses loss-landscape geometry to distinguish shortcut-driven from causal information, then disentangles corresponding parameter pathways and applies targeted unlearning updates.

  • Sharpness-Aware Partitioning: CUPID partitions the forget set by exploiting distinct loss-landscape geometry associated with bias-aligned and bias-conflicting samples.Bias-aligned samples cluster near flat regions, whereas bias-conflicting samples occupy sharper, more sensitive regions.
  • Sharpness-Aware Partitioning: Local sharpness is estimated from each sample’s loss change after a small, normalized gradient perturbation of the original parameters.Samples below the top-k% sharpness threshold form the bias-approximated set, while higher-sharpness samples form the causal-approximated set.
  • Causal Pathway Identification: Causal Pathway Identification separates parameters into causal and bias pathways using parameter magnitude and curvature-based saliency.The causal mask selects influential parameters using the diagonal Hessian and a percentile threshold; remaining parameters form the bias pathway.
  • Targeted Pathway Update: The targeted update aims to erase causal information while minimizing unintended changes to the bias pathway.CUPID projects the full forget-set gradient onto the causal gradient to obtain the causal component, treating the remaining orthogonal component as the bias gradient.
  • Targeted Pathway Update: Distinct gradients are routed to their corresponding pathways, with sharpness weighting the causal update so harder samples receive greater emphasis.The causal pathway receives the projected causal gradient, while the bias pathway receives the bias gradient, preventing shortcut unlearning.

Experiments

Experiments evaluate CUPID on biased datasets using forgetting, retention, robustness, privacy, ablation, sharpness-partition, and Grad-CAM analyses. CUPID achieves strong and balanced forgetting, including on unbiased test data, while its components and sharpness-based subsets support the intended causal-versus-bias intervention.

  • Evaluation Setup: Experiments use Waterbirds, BAR, and Biased NICO++, with Retain Accuracy, Forget Accuracy, △gap, WGA, and Membership Inference Attack metrics.Lower Forget Accuracy, △gap, and WGA indicate more effective or balanced unlearning, while MIA assesses privacy leakage.
  • Heavily Biased Training Set: 6.91% on Waterbirds, 7.70% on BAR, and 7.71% on Biased NICO++ are CUPID’s training-set Forget Accuracy values in a set with 99.5% bias-aligned samples.CUPID also achieves minimal △gap and lowest WGA without requiring access to the retain set.
  • Generalized Unlearning: 3.75% versus 30.26% is CUPID’s BAR Forget Accuracy compared with the next-best method on the unbiased test set.CUPID has the lowest Forget Accuracy, △gap, and WGA across datasets while maintaining high Retain Accuracy; its MIA scores are closer to Retrain than most alternatives.
  • Ablation Study: Ablations show that Sharpness-Aware Partitioning improves forgetting, Causal Pathway Identification restores high Retain Accuracy, and Targeted Pathway Update improves all forgetting-related metrics.The results support the contribution of all three CUPID components.
  • Sharpness-Aware Partitioning: At k = 5%, the causal-approximated set yields optimal performance despite being impure and enriched with hard bias-conflicting samples.Increasing k makes the threshold less restrictive and increases bias-aligned samples; easier samples may regularize the causal gradient.
  • Qualitative Comparison: Grad-CAM shows that CUPID avoids bias-related activation regions, whereas existing methods often continue activating on spurious features.The qualitative comparison covers Waterbirds, BAR, and Biased NICO++.

Conclusion

The paper formalizes shortcut unlearning as a failure mode in biased models and introduces CUPID as a surgical intervention. Experiments show superior forgetting while mitigating shortcut unlearning, with future work targeting entangled abstract concepts.

  • Conclusion: Shortcut unlearning describes models struggling to forget easily learned bias-aligned samples while erasing shortcut attributes instead of intended class information.This can paradoxically improve accuracy on bias-conflicting samples of the class being forgotten.
  • Conclusion: CUPID partitions samples, identifies neural pathways, and applies targeted updates to disentangle and erase causal information.The framework leverages loss-landscape geometry for this surgical intervention.
  • Conclusion: Experiments on standard biased datasets show that CUPID significantly outperforms existing methods while mitigating shortcut unlearning.The paper proposes adapting this approach to abstract concepts entangled with other knowledge.

Implementation Details

Implementation evaluates CUPID on three biased datasets using a common ResNet-50 experimental setup. Pretraining and unlearning use specified AdamW schedules and batch sizes on a single RTX 3090 GPU.

  • Datasets: The evaluation uses Waterbirds, BAR, and a constructed Biased NICO++ dataset, each pairing class labels with spurious backgrounds, places, or contexts.NICO++ supplies object and context annotations for constructing the biased version.
  • Training Setup: All experiments use PyTorch on one NVIDIA RTX 3090 GPU, with ResNet-50 models pretrained for 10 epochs using AdamW at learning rate 10^-4.Pretraining uses weight decay 10^-3, cosine annealing, and batch size 128; unlearning runs for one epoch at learning rate 10^-5 and batch size 64.

More Ablation Studies

The ablations show that CUPID’s effectiveness depends on both disentangled gradient routing and sharpness-based adaptive reweighting. Together, these components provide complementary coarse- and fine-grained control over forgetting.

  • Disentangled Gradient Update: Applying causal and bias gradients to their respective pathways outperforms applying either component alone.The combined update erases class-specific knowledge while neutralizing shortcut influence.
  • Disentangled Gradient Update: The standard forget-set gradient has high FA and △gap because its direction is dominated by shortcut features.
  • Disentangled Gradient Update: Applying only the causal gradient significantly improves FA, whereas applying only the bias gradient shows almost no improvement.The bias-only update leaves core class knowledge in the causal pathway intact.
  • Sharpness-Aware Reweighting: Sharpness-based adaptive reweighting is critical beyond initial sample partitioning, substantially improving forgetting effectiveness.Partitioning isolates a hard causal-approximated set, while reweighting applies stronger updates to samples in sharper loss-landscape regions.

Additional Qualitative Results

Additional Grad-CAM visualizations compare how unlearning methods activate features on forget-set samples. CUPID produces diffuse activation, like the Retrain reference, rather than focusing on shortcut or causal features.

  • Additional Qualitative Results: CUPID and Retrain show diffuse Grad-CAM activation with no focus on either feature, indicating successful erasure of the class concept.
  • Additional Qualitative Results: Most baseline methods continue activating either the spurious shortcut or the causal object on forget-set samples.
Loading 2602.21773v1…