Generative Perturbation Network for Universal Adversarial Attacks on Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are effective for electroencephalogram (EEG)-based brain-computer interface (BCI) classification.
• Adversarial examples, subtle input perturbations, can deceive even high-performing DNN models in BCI systems.
• These perturbations are often imperceptible to humans, posing a significant security risk.

Purpose of the Study:

• To propose an efficient generative model, the Generative Perturbation Network (GPN), for creating universal adversarial examples.
• To enable the generation of adversarial examples for both non-targeted and targeted attacks on BCI systems.
• To demonstrate the model's efficiency and effectiveness in generating robust and transferable adversarial perturbations.

Main Methods:

• Developed a Generative Perturbation Network (GPN) capable of generating universal adversarial examples.
• Extended the GPN to conditionally or simultaneously generate perturbations for diverse targets and victim models.
• Evaluated the performance of GPN-generated perturbations against previous adversarial attack methods.

Main Results:

• GPN-generated perturbations demonstrated superior performance compared to prior art in crafting signal-agnostic perturbations.
• The extended GPN significantly reduced generation time for signal-specific methods while maintaining comparable performance.
• The proposed method exhibited superior transferability of perturbations across different classification networks, indicating high generality.

Conclusions:

• The Generative Perturbation Network (GPN) offers an efficient and effective approach to generating universal adversarial examples for BCI systems.
• GPN enhances the security of BCI systems by creating robust perturbations that can deceive DNN models.
• The model's efficiency and transferability make it a valuable tool for BCI security research and development.