Physical Security Using Machine Learning to Detect Lock Picking at Traffic Cabinets

Abstract

Traffic systems are filled with essential traffic control equipment and can cause massive infrastructural damage and driver safety if hacked. We explore a machine learning method to detect real-time lock picking to thwart unauthorized access to the electronics. We gather accelerometer and gyroscopic data to train a decision tree model for detecting lock picking. Analysis reveals that a standard deviation feature for only two accelerometer axes is adequate for achieving robust performance. We deployed an real-time decision tree model to an offsite test cabinet that achieves an accuracy of over 95 %.