Detection of Acoustic Side‐Channel Attacks Based on Multidimensional Information Fusion

ABSTRACT

Accelerometer and audio sensors in industrial environments are vulnerable to acoustic side‐channel attacks, leading to false alarms in predictive maintenance systems, compounded by the issue of scarce fault sample data. This paper proposes a defense mechanism based on multidimensional information fusion to accurately determine whether sensors are under attack. The first layer of this defense mechanism dynamically adjusts confidence scores using object detection algorithms to monitor physical intrusions, thereby defending against noncontact attacks and enhancing the system's security capabilities. The second layer integrates accelerometer and audio data through multimodal data fusion, initially subjecting these data types to adversarial attacks to generate a new dataset for training fusion algorithms. This algorithm employs deep separable convolutional networks and residual networks to perform in‐depth analysis of accelerometer and audio data, feeding the extracted features into autoencoders (AEs) and convolutional attention mechanisms to derive relevant scores, thereby enhancing the detection capability against acoustic side‐channel attacks. Experimental results demonstrate that the proposed fusion algorithm outperforms other models, achieving an accuracy rate of 98.749%. This accuracy enables more precise identification of sensor attacks. Through this defense mechanism based on multidimensional information fusion, predictive maintenance systems can more reliably avoid false alarms, thereby enhancing system security and stability.