GenFighter: A Generative and Evolutionary Textual Attack Elimination Method

GenFighter is a defense method against NLP adversarial attacks, outperforming existing state-of-the-art defenses on the metrics of attacked accuracy and attack success rate. By learning and reasoning over the training classification distribution, it identifies potential malicious samples that deviate from the distribution, transforms them into instances semantically equivalent to the training data, and responds in a unified way with an ensemble method. The method also requires a large number of queries for each attack, making attacks harder to carry out in real-world scenarios.