The rapid proliferation of Internet of Things (IoT) devices has amplified the complexity of securing network infrastructures against sophisticated cyber-attacks. Traditional intrusion detection systems (IDS) struggle to generalize in dynamic IoT environments, where data is high-dimensional and often imbalanced. This study presents an advanced neural network-based intrusion detection framework tailored for IoT networks, focusing on optimizing detection accuracy while mitigating overfitting through dropout and L2 regularization techniques. Utilizing a comprehensive IoT intrusion dataset, we conducted extensive preprocessing, including feature scaling, outlier removal, and correlation analysis, to enhance model reliability and performance. Three model architectures were developed and evaluated: a baseline model without regularization, a dropout-only model, and a fully optimized model with both dropout and L2 regularization. Experimental results demonstrate that the fully optimized model achieved an accuracy of 87%, outperforming baseline models by effectively balancing recall and precision, especially for minority attack classes. Our findings underscore the critical role of regularization in neural network performance for IoT intrusion detection, suggesting that such models can provide robust defense mechanisms against evolving cybersecurity threats. Future research will explore ensemble methods, sequential architectures, and real-time data pipelines to refine IoT network security further.