The cybersecurity of autonomous vehicles is a rapidly evolving field, facing significant challenges as these vehicles become more integrated into modern transportation systems. Autonomous vehicles rely on complex networks of sensors, communication systems, and control units, all of which are vulnerable to cyber threats such as hacking, spoofing, cloning, and unauthorized access. With the development of sophisticated defense mechanisms and increasingly advanced cyber-attacks, one emerging solution in this context is physical fingerprinting, which involves using the unique physical characteristics of hardware component/electronic device to authenticate and monitor the integrity of critical cyber-physical systems. 

Physical fingerprinting provides an additional layer of security, helping to ensure that all hardware component/electronic device within the vehicle are legitimate and functioning as expected. As the adoption of autonomous vehicles grows, the implementation of robust cybersecurity measures, including physical fingerprinting, will be essential to safeguard these vehicles from evolving threats. 

In Cyber-Physical System (CPS), physical fingerprinting involves analyzing specific physical properties or responses of the system—such as electrical noise, electromagnetic emissions, power consumption patterns, acoustic signals, or thermal characteristics—that are inherently tied to the hardware and cannot be easily replicated or altered. In autonomous vehicles, physical fingerprinting can be used to uniquely identify each hardware component/electronic device within an autonomous vehicle. For instance, each Electronic Control Unit (ECU) within the vehicle have a distinct electrical noise or distortion, which can be extracted as a signature for device fingerprinting (DFP). These fingerprints help ensure that only legitimate and authenticated electronic control units (ECUs) are part of the vehicle’s network, preventing unauthorized tampering or substitution of critical parts.

Physical fingerprinting can also be applied to sensors, ensuring that data being fed into the vehicle’s control system is coming from legitimate sources and has not been spoofed by an attacker. By leveraging these unique physical unclonable fingerprints of the hardware, the system can detect anomalies that might indicate a security breach, such as hardware tampering or the introduction of counterfeit parts. This adds an additional layer of security to protect against cyber-attacks.

Two primary approaches for creating a framework that utilizes physical unclonable fingerprints are:

• Feature Extraction - This approach is based on extracting statistical features from the device's signal.

• Deep Learning - This approach is based on training a deep neural network on the raw signal.  

The integration of both feature based and deep learning based approaches to physical fingerprinting will be crucial for building robust cybersecurity defenses capable of addressing the growing complexity of potential threats.

The feature extraction approach involves deriving features from the output signal of an ADC (Analog to Digital Converter). The signal from the ECUs is passed through an ADC (such as those from Analog Devices, Digilent, or Microchip), where it is captured at a high sampling rate. The acquired signal, which contains physical characteristics, is then converted into digital form and stored in a CSV file. During the acquisition process, noise is introduced into the signal, and these noise patterns (fingerprints) are utilized to extract the signal's features. These noise patterns are unique for each individual device. These extracted features are subsequently used to train machine learning models. The trained models are then employed to predict the identity of the device (ECU). The trained models are then employed to predict the identity of the device (ECU). This method has achieved an accuracy rate of 99%.

Device fingerprinting, based on deep learning approach, is one of the cybersecurity measures that has its application for securing the cyber-physical systems including autonomous vehicles, robots etc. 

In this approach we present an end-to-end framework for electronic device fingerprint extraction, model training, and deploy device-specific models for real-time device identification/verification using an analog input signal. The output of an electronic device is used to compute distortion profile which is then transformed into spectrograms. These spectrograms are then applied at the input of deep learning architecture to learn the underlying device specific unique distortion profile - also referred as physical fingerprint.

The effectiveness of the proposed method is evaluated using rigorous set of experiments. Experimental evaluation is performed on 10 electronic devices of same make and model, running same code on them, and the data is collected through same data acquisition system. Experimental results indicate that the proposed framework achieves perfect device recognition even in real-time testing.