detect dataset shortcuts

Existing shortcut learning mitigation methods typically rely on fully annotated data, group-balanced validation sets, or training data that exhaustively covers all attribute–class combinations—conditions rarely met in real-world scenarios. This work proposes a novel approach that requires neither additional annotations nor balanced validation sets. By analyzing internal model representations, the method identifies a small set of samples exhibiting spurious correlations and locates critical neurons responsible for leveraging these misleading attributes, guided by the principle that such features should not inform predictions. Intermediate-layer regularization is then applied to these neurons. Requiring only a few spuriously correlated (false positive) samples, the approach effectively suppresses shortcut learning, significantly enhancing model robustness and preventing models from making correct predictions for incorrect reasons, thereby achieving effective mitigation under more realistic data conditions.

Can regularization effectively mitigate models’ reliance on spurious correlations (“shortcuts”) without compromising causal features? This paper systematically reveals its “double-edged sword” effect: mainstream methods—including DropBlock, RSC, and ERM+L2—often over-suppress causal features, degrading out-of-distribution (OOD) generalization. Through decomposition of generalization error and theoretical analysis of feature sensitivity, we first derive a formal criterion to distinguish causal from spurious features and rigorously establish necessary and sufficient conditions for effective regularization. We validate our theory via synthetic data and multiple OOD benchmarks (Colored MNIST, Waterbirds, CelebA), demonstrating that most regularizers significantly attenuate causal signals under standard hyperparameter settings. Building on this, we define a verifiable “safe regularization interval,” within which regularization enhances robustness without harming causal learning. Empirically, leveraging this interval improves OOD accuracy by up to 12.7% across benchmarks.

Shortcut learning in medical AI—where models erroneously rely on non-clinical imaging artifacts rather than pathologically relevant features—leads to spurious correlations and compromised clinical reliability. Method: We propose the first XAI-driven, semi-automated framework for shortcut identification and disentanglement, integrating gradient-based interpretability methods (Grad-CAM, Integrated Gradients), causal attribution analysis, adversarial data distillation, and model editing. It enables sample- and pixel-level bias localization and mitigation without expert re-annotation. The framework is architecture-agnostic (supporting CNNs and ViTs) and generalizable across multimodal medical data. Results: Evaluated on four medical datasets, our approach effectively identifies and eliminates artifact-induced spurious associations, significantly improving out-of-distribution robustness and clinical trustworthiness of VGG16, ResNet50, and ViT models.

This work addresses the problem of shortcut learning in binary black-box classification models caused by dataset bias. It proposes a novel post-hoc analysis framework that integrates interventional and observational perspectives, introducing linear mixed-effects models—used here for the first time—to diagnose bias in black-box classifiers. By decomposing the influence of training and test data on model scores, the method moves beyond conventional error-rate metrics to uncover the risk of models relying on spurious correlations. The approach effectively identifies and quantifies the impact of data bias on decision-making in voice anti-spoofing and speaker verification tasks, offering a new pathway toward building reliable and interpretable AI systems.

The layer-wise distribution mechanism of shortcuts (spurious correlations) in deep networks remains poorly understood, hindering principled mitigation strategies. To address this, we propose counterfactual inter-layer attribution to quantify each layer’s contribution to generalization degradation under clean versus biased data, conducting systematic analysis across VGG, ResNet, DeiT, and ConvNeXt on CIFAR-10, Waterbirds, and CelebA. We discover a cross-layer collaborative shortcut learning pattern: shallow layers predominantly encode spurious features, while deeper layers selectively forget core discriminative features from clean data. Leveraging this insight, we construct multi-dimensional perturbation axes for precise shortcut localization. Experiments reveal that shortcut effects permeate the entire network and exhibit strong dependence on both dataset and architecture, rendering generic mitigation strategies ineffective. Customized, architecture- and task-aware interventions are thus essential. This work establishes a novel paradigm for mechanistic modeling of shortcuts and targeted intervention.

This study addresses the performance degradation of deepfake audio detection models in real-world scenarios, often caused by reliance on dataset-specific shortcut features such as non-speech artifacts. The work introduces causal intervention into anti-spoofing diagnostics for the first time, proposing a directed graphical model-based intervention framework that formally distinguishes between shortcut learning and legitimate domain shifts. Through controlled acoustic perturbations—targeting non-speech segments, spectral characteristics, and energy profiles—and corpus-level distributional analysis, the authors systematically evaluate model sensitivity. Experiments on the ASVspoof dataset reveal that interventions on non-speech intervals lead to significant performance drops, identifying them as critical shortcut features and providing a clear direction for enhancing model robustness.

Real-world datasets often suffer from multiple issues, including label noise, feature corruption, and spurious correlations, yet existing methods struggle to simultaneously identify erroneous samples and their specific error types with high precision. This work proposes DeMix, a novel framework that, for the first time, leverages influence vectors to characterize how individual training samples affect predictions on a validation set. By formulating data debugging as a multi-label classification task and incorporating intervention-based learning to extract invariant diagnostic criteria for each error type, DeMix enables joint identification of both corrupted samples and their underlying error categories. Evaluated across 11 benchmark tasks, DeMix improves the F1 score for data debugging by 22.61% on average and boosts downstream model performance by 9.32% after data repair, substantially outperforming current state-of-the-art approaches.