Adversarial Attacks on Question Answering
The recent explosion of Question Answering datasets and models is pushing the boundaries of QA systems and making them widely used by the general public in virtual assistants or chatbots (Rogers et al., 2021). However, it is well-known that is possible to “trick” machine learning systems. For instance, Goodfellow et al. (2014) shows that a clear picture of a panda can be modified with some invisible noise for humans that makes the ML model classify it as a