Adversarial learning: the impact of statistical sample selection techniques on neural ensembles

Adversarial learning is a recently introduced term which refers to the machine learning process in the presence of an adversary whose main goal is to cause dysfunction to the learning machine. The key problem in adversarial learning is to determine when and how an adversary will launch its attacks. It is important to equip the deployed machine learning system with an appropriate defence strategy so that it can still perform adequately in an adversarial learning environment. In this paper we investigate artificial neural networks as the machine learning algorithm to operate in such an environment, owing to their ability to learn a complex and nonlinear function even with little prior knowledge about the underlying true function. Two types of adversarial attacks are investigated: targeted attacks, which are aimed at a specific group of instances, and random attacks, which are aimed at arbitrary instances. We hypothesise that a neural ensemble performs better than a single neural network in adversarial learning. We test this hypothesis using simulated adversarial attacks, based on artificial, UCI and spam data sets. The results demonstrate that an ensemble of neural networks trained on attacked data is more robust against both types of attack than a single network. While many papers have demonstrated that an ensemble of neural networks is more robust against noise than a single network, the significance of the current work lies in the fact that targeted attacks are not white noise.