Abstract:
We present a novel method for robustness training for ReLU-based deep neural networks. The method involves a decision tree search targeting the worst-case data points to generate adversarial examples. We combine the decision tree search method with robust optimisation to train a robust model while maintaining accuracy at comparably lower computational effort than SoA methods. The efficiency is obtained by focusing on small regions centred around the input that have significant potential to generate adversarial samples. We implemented the resulting method in the framework DTSRobust, which was evaluated against state-of-the-art defence methods on MNIST and CIFAR10 datasets. In experiments, DTSRobust achieved a 14.2% gain on efficiency against the state-of-the-art defence methods in MNIST and 10.3% of that in CIFAR10 while maintaining similar accuracy.
