Detecting the universal adversarial perturbations on high-density sEMG signals

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Myoelectric pattern recognition using Convolutional Neural Networks (CNNs) shows promise for upper limb neuroprosthetic control.
• CNN models are vulnerable to adversarial perturbations (UAP), significantly degrading performance in electromyography (EMG) signal classification.
• Existing CNN-based systems achieve >90% accuracy but drop below 20% under attack, posing a security risk.

Purpose of the Study:

• To address the lack of adversarial attack detection in myoelectric control systems.
• To propose a novel method for detecting adversarial attacks targeting EMG signals used in prosthetics.
• To enhance the security and reliability of muscle-computer interfaces.

Main Methods:

• A novel correlation feature based on Chebyshev distance between adjacent EMG channels was developed.
• The proposed detection framework was evaluated using two high-density EMG datasets.
• The method focuses on detecting Universal Adversarial Perturbations (UAP) in EMG signals.

Main Results:

• The proposed detection method achieved high detection rates of 91.39% and 93.87% on two separate datasets.
• The detection framework operated with a minimal latency of no more than 2 ms.
• The method effectively identified adversarial attacks on CNN-based myoelectric pattern recognition.

Conclusions:

• The developed correlation feature provides an effective means for detecting adversarial attacks in EMG signals.
• This detection framework offers early warning and defense capabilities against security threats in neuroprosthetic control.
• The findings contribute to improving the security and robustness of muscle-computer interfaces.