Password-Sniffing Acoustic Keylogger Using Machine Learning

Abstract

An acoustic keylogger is a form of side-channel attack where an attacker can extract and log what a victim typed on a keyboard by capturing the keystroke emanations and audio signals. Using a compromised smartphone microphone to capture keystroke audio recordings, this attack can be used to recognise confidential information such as passwords typed by a victim. In this paper, we demonstrate the training of a machine-learning model and develop a proof-of-concept password-sniffing acoustic keylogger. To this end, the input audio signals are initially converted to audio images by using the spectrogram representation of the Mel-Frequency Cepstral Coefficients (MFCC) method. MFCC applies the Fourier transform to yield the frequency content of the input signal. Then, a novel and less complex vision transformer-based approach namely ConvMixer is applied to the spectrogram images to detect the password according to its audio. In ConvMixer model group convolution and point convolution layers are located sequence for giving depth to the architecture. A novel key logger dataset is constructed by us and used in the experimental works. Various experiments are carried out and classification accuracy is applied for performance measurement tools. In the experimental works, an average 92.44% accuracy score is obtained with the proposed method. In addition, two pre-trained convolutional neural networks (CNN) models namely ResNet18 and VGG16 are used for comparison purposes. Comparisons show that the proposed approach has the potential for a compact acoustic keylogger system.

Keywords: Keylogger, Acoustic Keylogger, Password sniffing, Machine Learning, Convolutional Neural Networks, ConvMixer