Adversarial Attacks on Multimodal Language Models

Nowadays artificial intelligence and machine learning, in particular, are involved in almost all spheres of our life and gaining more and more popularity every day. So, it is very important to provide security vulnerability and robustness of deep learning models because even insignificant troubles with these aspects can lead to terrible consequences. Recently, multimodal language models became an object of interest for many researchers because such models give us an opportunity to solve a large amount of deep learning tasks. Adversarial attacks are among the most dangerous threats in such tasks. If we have a task, where input includes an image, such an attack means that we can fabricate samples by adding noise to them, which are indistinguishable from the correct samples for people, but at the same time our model will be broken and give us an incorrect output. Studying adversarial attacks on such models can improve their robustness and reduce vulnerability.