Unveiling Vulnerability of Self-Attention

Pre-trained language models (PLMs) are shown to be vulnerable to minor word changes, which poses a big threat to real-world systems. While previous studies directly focus on manipulating word inputs, they are limited by their means of generating adversarial samples, lacking generalization to versatile real-world attack. This paper studies the basic structure of transformer-based PLMs, the self-attention (SA) mechanism. (1) We propose a powerful perturbation technique textitHackAttend, which perturbs the attention scores within the SA matrices via meticulously crafted attention masks. We show that state-of-the-art PLMs fall into heavy vulnerability that minor attention perturbations (1%) can produce a very high attack success rate (98%). Our paper expands the conventional text attack of word perturbations to more general structural perturbations. (2) We introduce textitS-Attend, a novel smoothing technique that effectively makes SA robust via structural perturbations. We empirically demonstrate that this simple yet effective technique achieves robust performance on par with adversarial training when facing various text attackers. Code is publicly available at urlgithub.com/liongkj/HackAttend.