<p>Intrusion detection in Cyber–Physical-Systems (CPSs) is critical because of their tightly integrated nature where physical processes and software components work together through computer algorithms and internet connectivity. Industrial CPSs due to their distributed architecture and reliance on advanced technologies are getting increasingly vulnerable to cyberattacks, highlighting the need for robust, intelligent, and secure monitoring. For this, we propose <i>SecureNet-Fed: ConvNeXt_Bi-GRU_IDS</i>, a federated deep learning framework that ensures robust, scalable, and privacy-preserving intrusion detection. The architecture combines both spatial and temporal modeling through a hybrid pipeline integrated with federated learning and secure encryption. Spatial patterns are extracted using a ConvNeXt module, capable of identifying localized anomalies within CPS data. These features are then passed to a Bidirectional GRU (Bi-GRU) to capture sequential dependencies and evolving attack behaviors. A series of dense layers with dropout regularization follow, optionally incorporating attention mechanisms to prioritize critical temporal sequences. The final output layer uses Softmax activation for binary classification. To preserve data privacy, the proposed framework employs federated learning, allowing each CPS node to train locally while only sharing encrypted model weights. A comprehensive experimental setup was used to evaluate the proposed approach, incorporating a range of intrusion and normal data. Experiments demonstrate that SecureNet-Fed achieves improved detection accuracy, scalability, and resilience in real-world CPS environments.</p>