Synthetic EMG Based on Adversarial Style Transfer Can Effectively Attack Biometric-Based Personal Identification...

Area of Science:

• Biometrics and Cybersecurity
• Machine Learning and Signal Processing

Background:

• Electromyography (EMG) signals are utilized in biometric identification due to their complexity and person-specific characteristics.
• Current deep learning models for EMG-based identification are assumed to be secure against fabricated biological inputs.

Purpose of the Study:

• To investigate the vulnerability of EMG-based identification models to adversarial biological attacks.
• To develop a novel method for generating synthetic EMG signals that mimic an individual's unique signal style.

Main Methods:

• A generative adversarial network (GAN) with an individual-style transformer was employed.
• The method utilizes tiny leaked EMG data segments to extract personal signal style.
• Synthetic EMG signals with diverse content but consistent personal style were generated to attack deep EMG classifiers.

Main Results:

• The proposed attack achieved an average success rate of 99.41% in confusing identification models.
• An average success rate of 91.51% was achieved in manipulating identification models.
• The attack demonstrated high effectiveness against three well-recognized deep EMG classifiers using data from eighteen subjects.

Conclusions:

• Deep neural network-based EMG classifiers are vulnerable to sophisticated synthetic data attacks.
• The findings highlight the need to consider synthetic biological signal generation in the design of robust biometric identification systems.
• Ensuring personal identification security requires addressing potential vulnerabilities to adversarial synthetic data.