The growth of Internet of Things (IoT) networks has drastically improved attack surface, requiring intrusion detection systems (IDS) to ensure accuracy and privacy protection. To overcome these obstacles, we introduce a federated learning (FL) based IDSW that incorporates state-of-the-art preprocessing, smart feature optimization, and a new classification paradigm. During preprocessing, raw traffic data is subject to scrubbing at a vigorous level, normalization through scaling, and label encoding to maintain consistency and reduce noise in heterogeneous local datasets. For feature selection, the Hybrid Emperor Penguin–Quokka Swarm Optimization (HEPQSO) approach is utilized which balances exploitation and exploration to find the most discriminative features while addressing the dimensionality problem. These features are then utilized by a deep hybrid classifier where the Spike Gated Linear Unit (SGLU) facilitates non-linear representation learning, and a Vision Transformer-Temporal Convolutional Network (ViT–TCN) hybrid discovers both global spatial relationships and local temporal dynamics of intrusion patterns. Experimental analyses performed using benchmark intrusion detection datasets show that the system has a high performance compared to baseline models at all times, with an accuracy of 97.88%, precision of 96.16%, recall of 97.54%, F1-score of 97.39%, specificity of 97.62%, and MCC of 97.04%, thus proving its efficiency for safe IoT settings. This combination of state-of-the-art preprocessing, hybrid feature selection, and deep federated classification forms a robust IDS that can tackle the changing landscape of cyber intrusions.