Evaluating the Adversarial Robustness of Text Classifiers in Hyperdimensional Computing

Abstract

Hyperdimensional (HD) Computing leverages random high dimensional vectors (>10000 dimensions) known as hypervectors for data representation. This high dimensional feature representation is inherently redundant which results in increased robustness against noise and it also enables the use of a computationally simple operations for all vector functions. These two properties of hypervectors have led to energy efficient and fast learning capabilities in numerous Artificial Intelligence (AI) applications. Despite the increasing number of such AI HD applications, their susceptibility to adversarial attacks has not been explored, specifically in the text domain. To the best of our knowledge, this is the first research endeavour to evaluate the adversarial robustness of HD text classifiers and report on their vulnerability to such attacks. In this paper, we designed and developed n-grams based HD computing text classifiers for two primary applications of HD computing; language recognition and text classification, and then performed a set of character level and word level grey-box adversarial attacks, where an attacker’s goal is to mislead the target HD computing classifier to produce false prediction labels while keeping added perturbation noise as low as possible. Our results show that adversarial examples generated by the attacks can mislead the HD computing classifiers to produce incorrect prediction labels. However, HD computing classifiers show a higher degree of adversarial robustness in language recognition compared to text classification tasks. The robustness of HD computing classifiers against character-level attacks is significantly higher compared to word-level attacks and has the highest accuracy compared to deep learning-based classifiers. Finally, we evaluate the effectiveness of adversarial training as a possible defense strategy against adversarial attacks in HD computing text classifiers.