Detection of cyber attacks on electric vehicles withcan-bus communication protocol using deep learni̇ngmethods

As vehicles become more digital and connect to other systems, it has become crucial to make certain that in-car communication networks are safe from hackers. This is especially true for light commercial vehicles that don't have a lot of computing capacity. This thesis presents a lightweight Convolutional Neural Network (CNN)- based Intrusion Detection System (IDS) aimed at improving the security of Controller Area Network (CAN) bus connections. The main goal of the work is to find fake messages and bad modifications to CAN traffic using a small model that may be used on low-power embedded platforms. A hybrid dataset was created to build and test the suggested model. This dataset included both authentic and misleading attack instances. Using the Vector VN1630 interface and CANoe software, real CAN traffic was gathered from a fully functional electric vehicle that met the J1939 standard. The Raspberry Pi 5 and MCP2515 transceiver module were used to create the attack data by simulating attacks like spoofing, message insertion, denial-of-service (DoS), and replay attacks. These scenarios focused on several electronic control units (ECUs) when the car was both still and moving. The CNN model could find spatial patterns in each CAN frame by turning it into an 8×13 binary grayscale image. The suggested architecture has two convolutional layers, pooling operations, fully linked layers, and regularisation methods to stop overfitting. The model has fewer than 200,000 parameters, but it still did a great job at classifying things, getting 98.87% accuracy. The whole work, from generating the data set to training an algorithm, had been completed under real-world situations. The model was able to learn subtle patterns that typical rule-based systems often miss by keeping the bit-level structure and time dynamics of CAN communications. Visualisations of the model's learned representations through PCA and t-SNE validate its capacity to distinctly differentiate between authentic and malicious samples. This study presents a novel deep learning approach for CAN Bus cybersecurity. The system that comes out of this is a viable and scalable way to identify intrusions in real time in modern automobiles, especially on platforms with limited processing capacity.