<p>The rapid evolution of the Internet of Things (IoT) has revolutionized data transmission and processing across diverse domains, enabling seamless automation without human intervention. Despite these advancements, IoT networks remain vulnerable to a wide range of security threats. Anomaly-based intrusion detection systems (IDSs) serve as a critical line of defence against such threats. In this research, we propose a novel AI-driven and optimized deep learning-based anomaly detection framework that integrates advanced augmentation, feature extraction, classification, and mitigation strategies to enhance intrusion detection performance significantly. To address the class imbalance, a modified conditional tabular generative adversarial network (MCTGAN) is employed to synthesize diverse and realistic minority-class samples, improving robustness and generalization. For feature extraction, we introduce a novel Squeeze-ViT framework that synergizes SqueezeNet and Vision Transformer architectures, capturing both local and global patterns in network traffic. Intrusion classification is performed using an optimized Fuzzy Capsule Network (Fuzzy-CapsNet), where hyperparameters are tuned using the Orangutan optimization algorithm (OOA) to achieve precise attack categorization. Most notably, an AI-driven mitigation mechanism is integrated through an improved deep Q-Network (IDQN), which automatically blocks IP addresses based on the severity of the detected threat. This severity-aware, autonomous response strategy introduces an intelligent and adaptive layer of network defence, setting a new standard for real-time IoT security solutions. Extensive evaluations using benchmark datasets validate the proposed approach, exhibiting high performance, and real-time mitigation efficiency.</p>