Autonomous vehicles depend on deep learning models for perception and wireless communication protocols for control and sensor data exchange, both of which present critical attack surfaces. This study investigates the security vulnerabilities of self-driving systems by evaluating adversarial attacks on perception models and network-based attacks on communication protocols. For perception, we assess the effectiveness of Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks on a Convolutional Neural Network (CNN) trained for the JetRacer platform. Using the CIFAR-10 and KITTI datasets, our findings show that PGD reduces the accuracy of the model by 9 to 13%, while targeted attacks on KITTI yield a 47.16% misclassification success rate. In parallel, we analyze the impact of man-in-the-middle (MitM), packet injection, and denial-of-service (DoS) attacks on autonomous vehicle control using JetRacer and Yahboom platforms. Using tools such as Bettercap and Wireshark, we demonstrate that network packet manipulation leads to lane deviation, control instability, and potential collisions. Together, these results highlight the multifaceted security risks in autonomous driving systems and emphasize the urgent need for robust defense mechanisms across both perception and communication layers.