Adversarially Optimized Mixup for Robust Classification

• DOI: 10.48550/arxiv.2103.11589
• Publication Date: 2021-01-01
• Authors: Jason Bunk; Srinjoy Chattopad...; B. S. Manjunath; Shivkumar Chandra...
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; Machine Learning (cs.LG)

Mixup is a procedure for data augmentation that trains networks to make smoothly interpolated predictions between datapoints. Adversarial training strong form of optimizes worst-case in compact space around each data-point, resulting neural much more robust predictions. In this paper, we bring these ideas together by adversarially probing the datapoints, using projected gradient descent (PGD). The fundamental approach work leverage backpropagation through mixup interpolation during optimize places where network makes unsmooth and incongruous Additionally, also explore several modifications nuances, like optimization ratio geometrical label assignment, discuss their impact on enhancing robustness. Through ideas, have been able train robustly generalize better; experiments CIFAR-10 CIFAR-100 demonstrate consistent improvements accuracy against adversaries, including recent ensemble attack AutoAttack. Our source code would be released reproducibility.