A New Deep Learning Pipeline for Acoustic Attack on Keyboards

The increasing reliance on services based on recent Artificial Intelligence advancements has elevated concerns about security vulnerabilities, leading to the exploration of novel attack vectors such as keystroke acoustic attacks on keyboards. This research delves into a deep learning approach for such attacks, which exploits acoustic emissions produced during typing to infer sensitive information. Traditional methods of keystroke acoustic attacks have relied on hand-engineered features and shallow classifiers, often failing to capture the intricate patterns within the acoustic data. In contrast, deep learning models have demonstrated remarkable capabilities in learning intricate patterns from complex data sources. We propose the exploitation of a Temporal Convolutional Network (TCN) to process acoustic signals, providing a more sophisticated and adaptive approach for keystroke acoustic attack analysis. The employed deep learning model showcases superior performance in multiple dimensions achieving a peak validation accuracy of 98.3% for keystrokes recorded by phone, and 93.05% for keystrokes recorded via Zoom, obtaining the best performances with respect the related prior art.