Building Adversary-Resistant Deep Neural Networks without Security through Obscurity

Deep neural networks (DNNs) have proven to be quite effective in a vast array of machine learning tasks, with recent examples in cyber security and autonomous vehicles. Despite the superior performance of DNNs in these applications, it has been recently shown that these models are susceptible to a particular type of attack that exploits a fundamental flaw in their design. This attack consists of generating particular synthetic examples referred to as adversarial samples. These samples are constructed by slightly manipulating real datapoints in order to “fool” the original DNN model, forcing it to mis-classify previously correctly classified samples with high confidence. Addressing this flaw in the model is essential if DNNs are to be used in critical applications such as those in cyber security. Previous work has provided various defense mechanisms by either augmenting the training set or enhancing model complexity. However, after a thorough analysis, we discover that DNNs protected by these defense mechanisms are still susceptible to adversarial samples, indicating that there are no theoretical guarantees of resistance provided by these mechanisms. To the best of our knowledge, we are the first to investigate this issue shared across previous research work and to propose a unifying framework for protecting DNN models by integrating a data transformation module with the DNN. More importantly, we provide a theoretical guarantee for protection under our proposed framework. We evaluate our method and several other existing solutions on both MNIST, CIFAR-10, and a malware dataset, to demonstrate the generality of our proposed method and its potential for handling cyber security applications. The results show that our framework provides better resistance compared to state-of-art solutions while experiencing negligible degradation in accuracy.