Brain-Computer Interface Security: Vulnerabilities of the Capsule Network Model

Recently, deep learning approaches have been widely applied to improve the classification performance of motor imagery-based brain-computer interfaces (BCI). In this scope, the capsule network (CapsNet) model emerges as a power-ful tool that implicitly learns various features, thereby achieving more reliable performance than traditional CNN approaches. While these advancements have significantly enhanced the accuracy and efficiency of BCI systems, the security aspect of these deep learning models has become a growing concern given the potential consequences of unauthorized access or manipulation of neural data. This paper extends the exploration of security concerns in BCI by introducing CapsNets, a relatively novel deep learning architecture, and investigates their robustness in the face of adversarial attacks. Our analysis is divided into two parts: first, we introduce a high-performance of CapsNet model for classifying brain signals based on EEG; second, we conduct a series of experiments to comprehensively assess attack strategies on EEG-based CapsNets. Results obtained from the well-known BCI competition 2b dataset showcase the efficacy of CapsNet in BCI tasks. However, our findings also reveal the vulnerability of these models to adversarial manipulation.