Behavioural Biometric Identification Using MIDI Keyboard Performance Signatures

Abstract

Musical performances encode distinctive behavioral signatures which are unique to each individual per-former, including subtle variations in timing, dynamics, and articulation. This study shows that these signatures, captured in MIDI data, can serve as reliable behavioral biometrics for performer identification and authorization. Through a newly developed Behavioural Biometric Authorization System (BBAS), keyboard performance MIDI data was recorded to extract structured sets of behavioral features (including inter-onset intervals, velocity, articulation, and polyphony). Three distinct machine learning classification models were trained on said features. Initial cross-validation techniques yielded classification accuracies between 94.87% and 92.31%, however, this precision was highly inflated due to limited dataset size of 5 users only. When evaluating data from new performances, the models were able to correctly identify musicians with a mean confidence of approximately 80%, and a minimum confidence of 67.47%