Sparse Adversarial Perturbation-Driven Scalable Coreset Optimization

Tushar Shinde, Manasa Madabhushi

Keywords: Data-efficient learning, coreset selection, sparse adversarial perturbations, adversarial robustness, low-data regimes, scalable optimization, distribution shift, corrupted data

Abstract: Efficient training of deep neural networks under limited data requires selecting informative subsets, or coresets, that preserve performance. Existing methods, such as DeepCore, rely on heuristics like uncertainty or gradient diversity, often overlooking adversarial vulnerabilities, which can degrade robustness under distribution shifts, corruptions, or manipulations. We introduce a unified \textbf{Adversarial Sensitivity Scoring} framework with three ranking strategies: Inverse Sensitivity and Entropy Fusion (ISEF), Fast Gradient Sign Method with Composite Scoring (FGSM-CS), and Perturbation Sensitivity Scoring (PSS), which leverage sparse adversarial perturbations to prioritize samples near decision boundaries. By applying single-step Sparse FGSM attacks, our methods reveal sample sensitivities with minimal computational cost. On CIFAR-10 with ResNet-18, our approaches outperform DeepFool by up to 15\% in extremely sparse (1\%) and low-data (10\%) regimes, while matching top DeepCore methods in moderate-data settings. Bottom-ranked variants excel in sparse regimes by retaining perturbation-resilient samples, whereas top-ranked variants dominate beyond 20--30\% data. These gains, achieved via efficient single-step gradients, position our framework as a scalable, deployable bridge between coreset selection and adversarial robustness, advancing data-efficient learning.

Submission Number: 120