Predicting the Efficiency of Adversarial Instancesfor Language of Graphs

A method for predicting adversarial instances in a language includes producing graphs for each sentence that capture the structure. The graph comprises nodes and edges, a node label represents a perturbed word, and the edge labeled notes possible typos of perturbations. Language labels are predicted for the unlabeled movie review which involves propagating labels through the graph. The adversarial instances appear nearly identical to the given input, but their equivalent prediction can be misinterpreted. In the past few years, assessing adversarial instances has become a decorum to estimate the robustness of deep learning models. While most of the literature emphasizes continuous features (i.e., image). However, some deep learning models also became successful on discrete features, and thus demand a model to understand underlying vulnerability and robustness. In this study, all the one-edit perturbations of a specific word in a review sentence of the Stanford Sentiment Treebank textual dataset which swaps the sentiment from positive to negative are analysed. Classifier's accuracy is also observed by inducing different typos and validated results indicate that graph structure alone is sufficient to predict whether an instance can fool the classifier or not.