On the Vulnerability of CNN Classifiers in EEG-Based BCIs

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models, particularly Convolutional Neural Networks (CNNs), excel in electroencephalogram (EEG)-based brain-computer interface (BCI) applications.
• However, deep learning models are susceptible to adversarial examples—subtly altered data that can cause misclassification.

Purpose of the Study:

• To investigate the vulnerability of CNN classifiers used in EEG-based BCIs to adversarial attacks.
• To propose and evaluate an unsupervised method for generating adversarial examples against BCI classifiers.
• To assess the transferability of these adversarial examples.

Main Methods:

• An Unsupervised Fast Gradient Sign Method (UFGSM) was developed to generate adversarial examples.
• The UFGSM was used to attack three popular CNN classifiers commonly employed in BCIs.
• The transferability of adversarial examples was tested across different models and datasets.

Main Results:

• The proposed UFGSM effectively attacked the targeted CNN classifiers in EEG-based BCIs.
• Adversarial examples demonstrated transferability, meaning attacks can succeed without prior knowledge of the target model's specifics.
• This study provides the first evidence of CNN classifier vulnerability to adversarial attacks in the context of EEG-based BCIs.

Conclusions:

• CNN classifiers in EEG-based BCIs are vulnerable to adversarial attacks, including those generated by unsupervised methods.
• The transferability of adversarial examples poses a significant security risk to BCI systems.
• Further research is needed to develop robust defenses and enhance the security of BCI technologies.