The rapid development of the Internet of Things (IoT) has brought significant security challenges. Due to the wide variety of devices, limited resources, and insufficient security measures, IoT devices are vulnerable to network attacks. Traditional intrusion detection methods struggle to handle high-dimensional, heterogeneous, and real-time traffic data, and face challenges from diverse attack patterns. To address these issues, this paper proposes a self-supervised and robust deep learning-based intrusion detection model—CAT-AE2. To strengthen the model’s feature extraction, we incorporate the Convolutional Block Attention Module (CBAM), which integrates channel and spatial attention to boost classification performance. We also propose an enhanced Temporal Convolutional Network (TCN) to capture long-term dependencies and complex temporal features in sequential data. Additionally, we have improved the discriminator design by incorporating convolution blocks, inter-mediate layers, and global average pooling layers, enabling the model to extract traffic features at different scales, reduce the risk of overfitting, and enhance stability under varying network conditions. Experimental results show that CAT-AE2 performs excellently on the NSL-KDD, UNSW-NB15, and CIC-IDS 2017 datasets, particularly under conditions of small samples and fluctuating traffic. The model effectively meets the demand for efficient, real-time, and robust detection in IoT environments.