Frequency Domain Channel-Wise Attack to CNN Classifiers in Motor Imagery Brain-Computer Interfaces.

Area of Science:

• Deep Learning
• Brain-Computer Interfaces
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are widely used in motor imagery brain-computer interfaces (MIBCI).
• Existing methods to evaluate CNN vulnerability in MIBCI primarily use temporal perturbations.
• These methods often require detailed knowledge of the victim model.

Purpose of the Study:

• To propose a novel attack approach for MIBCI systems based on frequency domain perturbations.
• To evaluate the effectiveness of this new approach against common CNN classifiers.
• To demonstrate a more flexible attack strategy that does not require detailed model information.

Main Methods:

• Developed a frequency domain channel-wise attack (FDCA) generating perturbations in the frequency domain.
• Utilized a differential evolution algorithm for optimization in a black-box scenario.
• Applied FDCA to attack three major CNN classifiers on four public MI benchmarks.

Main Results:

• FDCA achieved a significantly higher success rate in attacking CNN classifiers compared to existing methods and baselines.
• The proposed frequency domain perturbations proved effective even in a black-box setting.
• Demonstrated the efficacy of FDCA across multiple public MI datasets.

Conclusions:

• Frequency domain perturbations offer a highly competitive and effective strategy for attacking MIBCI systems utilizing CNN models.
• FDCA provides a flexible and powerful alternative to gradient-based attacks, offering enhanced model protection.
• This research advances the understanding of MIBCI vulnerability and security in black-box scenarios.