In this paper, we investigate the two most popular families of deep neural architectures (i.e., ResNets and InceptionNets) for the autonomous driving task of steering angle prediction. To ensure a comprehensive comparison, we conducted experiments on the Kaggle SAP dataset and custom dataset and carefully examined a range of different model sizes within both the ResNet and InceptionNet families. Our derived models can achieve state-of-the-art results in terms of steering angle MSE. In addition to this analysis, we introduced the attention mechanism to enhance steering angle prediction. This attention mechanism facilitated an in-depth exploration of the model's selective focus on essential elements within the input data. Furthermore, recognizing the importance of security and robustness in autonomous driving assessed the resilience of our models to adversarial attacks.