Universal adversarial perturbations for CNN classifiers in EEG-based BCIs.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Convolutional Neural Network (CNN) classifiers are widely used in electroencephalogram (EEG)-based brain-computer interfaces (BCIs).
• CNN models are susceptible to universal adversarial perturbations (UAPs), which are small, example-independent additions that can significantly impair model performance.
• UAPs pose a potential security threat to the reliability of BCI systems.

Purpose of the Study:

• To propose a novel total loss minimization (TLM) approach for generating UAPs specifically for EEG-based BCIs.
• To evaluate the effectiveness of the TLM approach against common CNN classifiers used in BCIs.
• To investigate the transferability of UAPs within EEG-based BCI systems.

Main Methods:

• Development of a total loss minimization (TLM) technique to generate UAPs.
• Application of TLM to create UAPs for EEG signals.
• Testing UAP effectiveness on three popular CNN classifiers used in EEG-BCIs.
• Evaluation of both target and non-target attack scenarios.
• Assessment of UAP transferability across different CNN models in BCI contexts.

Main Results:

• The proposed TLM approach effectively generated UAPs that degraded the performance of CNN classifiers in EEG-based BCIs.
• Demonstrated effectiveness for both target and non-target adversarial attacks.
• Confirmed the transferability of UAPs, meaning perturbations generated for one model can affect others.
• Experimental validation on three widely-used CNN architectures.

Conclusions:

• This research presents the first study on UAPs targeting CNN classifiers in EEG-based BCIs.
• The TLM method provides an effective means to generate potent UAPs for EEG-BCIs.
• The ease of UAP construction and real-time attack capability reveal a critical security vulnerability in current BCI technology.