Replication Data for: Unveiling Illusionary Robust Features: A Novel Approach for Adversarial Defenses in Deep Neural Networks

In this paper, we used three widely used benchmark datasets: CIFAR-10, MNIST, and CINIC-10. CIFAR-10 contains 60,000 color images (50,000 for training and 10,000 for testing) across 10 classes for object recognition tasks. MNIST consists of 70,000 grayscale images of handwritten digits, commonly used for image classification research. CINIC-10 extends CIFAR-10 with a selection of ImageNet images, totaling 270,000 images across the same categories, supporting scalability and robustness evaluations. Together, these datasets provide a foundation for research in computer vision, neural networks, and transfer learning.

In the paper, "Unveiling Illusionary Robust Features," we applied our purified robustification method to CIFAR-10 and CINIC-10. The resulting purely robustified CIFAR-10 dataset is available here: (train_data available in "https://drive.google.com/file/d/1kUlo2mHeTaA-fdEonaeK-rts7bhm4l5y/view?usp=sharing", and train_labels are available in "https://drive.google.com/file/d/1p7msdDBnbg4Jz5NGBG-roNjczR7QvkVd/view?usp=sharing").