Image Shortcut Squeezing is a machine learning data preprocessing defense that applies image compression techniques to neutralize perturbative availability poisons in training data. Perturbative availability poisons introduce subtle, imperceptible alterations to images that act as artificial shortcuts, which trick neural networks into memorizing noise rather than learning genuine, generalizable visual features. By compressing training images before model optimization, Image Shortcut Squeezing removes or disrupts these fragile shortcut perturbations, effectively restoring the training utility of protected or poisoned datasets with minimal computational overhead compared to retraining-based defenses such as adversarial training.
