Audio Signal Analysis and Classification of Bluetooth Vulnerabilities Using Machine Learning Techniques

Bluetooth has become one of the most popular short-range wireless communication standards, replacing traditional cable connections. However, recent discoveries have shed light on potential hazards associated with wireless communication. In 2020, the Braktooth vulnerability was identified as a means to assess the susceptibility of Bluetooth-enabled system-on-chips, modules, and end products to attacks. Given the prevalent use of Bluetooth in audio streaming applications, it is crucial to evaluate the impact of these attacks on audio signals. Since the effects of each vulnerability may not be discernible by listening alone, a comprehensive analysis is required. This research focuses on the classification of Bluetooth vulnerabilities, representing a critical step toward enhancing Bluetooth device security. The Braktooth exploiter kit was employed to implement various Braktooth attacks on the NXP BT system-on-chip on the iMX8xx NXP platform. The audio files resulting from the exploited attacks were recorded and exported using the Ellisys Vanguard Bluetooth Analysis System. A total of 1,660 audio signals were subjected to different Braktooth attacks, from which 89 numerical features were extracted. Three primary feature arrays, namely chromagram, spectrogram, and MFCC, were utilized to derive these numerical features. The collected numerical features were then employed to train a classical machine learning model for the purpose of attack classification. Among various models evaluated, the multi-layer perceptron model was selected, exhibiting an impressive accuracy of 98.04% on the training dataset and 57.53% on the testing dataset when appropriate modifications were made to the hyperparameters. Consequently, our proposed approach successfully establishes a robust Bluetooth vulnerability classification system.