Misleading Sentiment Analysis: Generating Adversarial Texts by the Ensemble Word Addition Algorithm

Deep neural networks are vulnerable to the adversarial examples that are generated by adding small perturbations to the original inputs. Similarly, traditional machine learning models such as logistic regression and support vector machine are also threatened by such carefully generated adversarial examples. In this paper, we study the adversarial texts that could mislead the sentiment analysis of traditional machine learning models. We present the Ensemble Word Addition (EWA) algorithm, which filters out a small number of words that have large attack ability and these words are added at the end of the original text. By extensive experiments, we show that the generated adversarial texts could fool some traditional machine learning models with a very high confidence, but they cannot affect human's judgment. The experimental results show that the proposed algorithm has a powerful attack ability to misleading the sentiment analysis, specifically the attack success rate could reach above 95% by only adding 4.6% words.