EEG-Based Brain-Computer Interfaces are Vulnerable to Backdoor Attacks.

Area of Science:

• Neuroscience
• Computer Science
• Cybersecurity

Background:

• Electroencephalogram (EEG) based brain-computer interfaces (BCIs) have advanced due to improved brain understanding and machine learning (ML) for signal decoding.
• ML algorithms in BCIs are susceptible to adversarial attacks, posing security risks.
• Existing adversarial attack methods can be complex to implement.

Purpose of the Study:

• To propose a novel, simplified method for adversarial poisoning attacks on EEG-based BCIs.
• To demonstrate the creation of backdoors in ML models by injecting poisoned training data.
• To highlight the critical security vulnerabilities in current EEG-BCI systems.

Main Methods:

• Introduction of a narrow period pulse for poisoning attacks in EEG-BCI training datasets.
• Development of a backdoor key that does not require synchronization with EEG trials for activation.
• Testing the effectiveness and robustness of the proposed backdoor attack method.

Main Results:

• The proposed narrow period pulse poisoning attack successfully creates backdoors in ML models for EEG-BCIs.
• Backdoor keys were shown to be effective without synchronization, simplifying attack implementation.
• The attack demonstrates significant robustness, posing a critical security threat.

Conclusions:

• The novel backdoor attack method presents a significant and easily implementable security concern for EEG-based BCIs.
• Urgent attention and development of robust defenses are required to mitigate these vulnerabilities.
• The findings underscore the need for enhanced security protocols in BCI research and development.