Abstract Cyber Physical Systems that support critical infrastructure are increasingly exposed to sophisticated cyber threats due to expanding connectivity and system complexity, while conventional anomaly detection approaches remain limited in scalability, adaptability to emerging attack patterns, and computational efficiency across heterogeneous environments. This study introduces a generalizable deep learning framework for anomaly detection in diverse CPS domains using a streamlined Deep Neural Network architecture that avoids complex ensemble designs while maintaining high performance. The framework integrates a structured preprocessing pipeline including normalization, categorical encoding, SMOTE based class balancing, and embedded feature selection to transform raw network traffic into discriminative inputs. Evaluation on the EdgeIIoT2023 and CICIoT2023 datasets, representing industrial and consumer CPS environments, demonstrates strong cross dataset robustness rather than universal domain independence, achieving up to 99.77% accuracy and an AUC of 0.9997. Compared with contemporary boosting and hybrid models, the proposed approach delivers competitive classification capability with lower computational complexity, supporting scalable real time deployment in mission critical industrial systems and resource constrained IoT networks while establishing a foundation for adaptable and efficient CPS security solutions.
Belew et al. (2026) studied this question.