Automobile Driver Fingerprinting

In advanced car systems, information passes through many units (ECUs) using CAN-Bus protocol. It is possible to extract wide range of data from many sensors, giving information about the driver, its state etc. This work focuses on data processing, after data extraction from a car. Using the data to identify the driver, out of 10 drivers, similarly to previous project (done by another team). In contrary to previous project, different approach is taken to feature extraction. This approach uses non-linear methods of dimension reduction, such as kernel-PCA and Diffusion Map.
The extracted data from a car is given in time-series format, that goes through pre-processing for data filtering, after then samples arranging according to hyper-parameters and using Diffusion Map algorithm resulting samples for each driver, classified by Random Forest. Using previous projects conclusions, this project uses Random Forest classifier. The results are not satisfying: most samples were not identified to the appropriate driver. Therefore the assumption that Diffusion Map is based on the data spreading on some non-linear space, is incorrect for the collected data about the drivers.
Reviewing all the results of the solutions brought through the project, the approach of extraction statistical characteristics gives better results.

Automobile Driver Fingerprinting