A Keystroke Audio-based Password Prediction Attack using Deep Learning

Protecting personal information and sensitive data has recently become a significant issue that has piqued the interest of security researchers. To access confidential information, users frequently input passwords via the keyboard. This current password input method will expose numerous risks and vulnerabilities when artificial intelligence, particularly deep neural networks (DNN), makes significant strides in audio analysis tasks. As a result, password prediction attacks based on DNN models' keyboard acoustic analysis are foretold to pose serious security risks to the system. This research proposes using deep learning techniques to process and analyze keyboard sounds to identify leaked information. We create a custom keystroke dataset to increase the system's performance during the training and evaluation phases of the deep learning model. An effective technique that utilizes the cut function has been developed to enhance the number of samples in the processing stage. Experimental results show that our proposed method for predicting characters and character sequences based on keystroke audio reaches an overall accuracy of up to 94.6%, which is suitable for practical applications.