Good, Better, Best: Textual Distractors Generation for Multi-Choice VQA via Policy Gradient

An increasing amount of studies have investigated the decision-making process of VQA models. Many of these studies focus on the reason behind the correct answer chosen by a model. Yet, the reason why the distracting answer chose by a model has rarely been studied. To this end, we introduce a novel task called \textit{textual Distractors Generation for VQA} (DG-VQA) that explaining the decision boundaries of existing VQA models. The goal of DG-VQA is to generate the most confusing set of textual distractors in multi-choice VQA tasks which expose the vulnerability of existing models (i.e. to generate distractors that lure existing models to fail). We show that DG-VQA can be formulated as a Markov Decision Process, and present a reinforcement learning solution to come up with distractors in an unsupervised manner. The solution addresses the lack of large annotated corpus issues in previous distractor generation methods. Our proposed model receives reward signals from fully-trained multi-choice VQA models and updates its parameters via policy gradient. The empirical results show that the generated textual distractors can successfully attack several popular VQA models with an average $20\%$ accuracy drop from $64\%$. Furthermore, we conduct adversarial training to improve the robustness of VQA models by incorporating the generated distractors. Empirical results validate the effectiveness of adversarial training by showing a performance improvement of $27\%$ for the multi-choice VQA task.