Discrete Adversarial Attacks and Submodular Optimization with Applications to Text Classification

Adversarial examples are carefully constructed modifications to an input that change the output of a classifier but are imperceptible to humans. these successful attacks for continuous data (such as image and audio), generating adversarial examples for discrete structures such as text proven significantly more challenging. In this paper we formulate the with discrete input on a set function as an optimization task. We prove this set function is submodular for some popular neural network text under simplifying assumption. This finding guarantees a 1-1/e factor for attacks that use the greedy algorithm. Meanwhile, we how to use the gradient of the attacked classifier to guide the greedy. Empirical studies with our proposed optimization scheme show improved attack ability and efficiency, on three different text tasks over various baselines. We also use a joint sentence and paraphrasing technique to maintain the original semantics and syntax of text. This is validated by a human subject evaluation in subjective metrics the quality and semantic coherence of our generated adversarial text.