Detecting Audio Adversarial Examples in Automatic Speech Recognition Systems Using Decision Boundary Patterns.

Area of Science:

• Computer Science
• Machine Learning
• Cybersecurity

Background:

• Automatic Speech Recognition (ASR) systems are widely used but vulnerable to audio Adversarial Examples (AEs).
• AEs manipulate ASR models, causing incorrect transcriptions and posing a significant security risk.
• The intrinsic differences between AEs and benign audio, particularly concerning decision boundaries, remain underexplored.

Purpose of the Study:

• To investigate the fundamental properties of audio Adversarial Examples (AEs) and benign audio.
• To identify distinct machine learning decision boundary patterns associated with AEs.
• To develop a method for detecting audio AEs based on these pattern differences.

Main Methods:

• Analysis of machine learning decision boundary patterns for audio AEs and benign audio.
• Application of dimensionality-reduction techniques to visualize these patterns.
• Utilizing anomaly detection methods for AE identification.

Main Results:

• Machine learning decision boundary patterns for audio AEs and benign audio are fundamentally different.
• These distinct patterns are visually distinguishable in two-dimensional (2D) space using dimensionality reduction.
• The identified patterns facilitate the detection of audio AEs.

Conclusions:

• The distinct decision boundary patterns offer a novel approach for identifying audio Adversarial Examples.
• Dimensionality reduction and anomaly detection can be effectively employed to enhance ASR system security.
• Further research into AE properties can lead to more robust defense mechanisms for speech recognition systems.