Fraud’s Bargain Attacks to Textual Classifiers via Metropolis-Hasting Sampling (Student Abstract)

Recent studies on adversarial examples expose vulnerabilities of natural language processing (NLP) models. Existing techniques for generating are typically driven by deterministic heuristic rules that agnostic to the optimal examples, a strategy often results in attack failures. To this end, research proposes Fraud's Bargain Attack (FBA), which utilizes novel randomization mechanism enlarge searching space and enables high-quality be generated with high probabilities. FBA applies Metropolis-Hasting algorithm enhance selection from all candidates proposed customized Word Manipulation Process (WMP). WMP perturbs one word at time via insertion, removal, or substitution contextual-aware manner. Extensive experiments demonstrate outperforms baselines terms success rate imperceptibility.