Visual Analysis on Adversarial Images

Adversarial Machine Learning (ML) is one of the biggest cyber threats to deep neural networks (DNN). Adversarial samples are crafted or created by an attacker to mislead a DNN model in making a decision. On the presence of adversarial samples in a dataset, an interest is to trace their patterns before designing an effective defense algorithm. In this paper, we propose a hybrid method, i.e. Convo t-SNE where convolution neural network (CNN) is used to extract the features of images and t-distributed Stochastic Neighborhood Embedding (t-SNE) is applied to reduce dimensions so that adversarial attack patterns can be visualized. In our work, three attack methods are applied on three public datasets (i.e., MNIST, CIFAR10 and CIFAR100). The graphical results indicate clearly the presence of attack patterns using Convo t-SNE. The proposed method is then hybrid with k-mean to further evaluate its effectiveness. Two evaluation metrics i.e. classification accuracy and a clustering normalized mutual information (ClustNMI, a metric to compare between two clustering algorithms, the higher the better) are used. The results are encouraging where Convo t-SNE improves the clustering algorithms with a higher classification accuracy and better ClustNMI than t-SNE.