A cyber-physical system (CPS) combines several interconnection physical methods networking units, and computing resources, and monitoring the applications and processes of the computing systems. Interconnected cyber and physical worlds start threatening security problems, particularly with the enhanced difficulty of network communications. Although efforts to address these problems, it can be complex to analyze and detect cyber-physical attacks in difficult CPS. Deep learning (DL)-driven techniques are implemented to investigate cyber-physical security systems. Therefore, this study presents a chaotic Harris Hawks optimization-based feature selection with attention-based DL (CHHOFS-ADL) technique for intrusion detection in the CPS platform. The drive of the CHHOFS-ADL technique is to ensure safety in the CPS platform through the intrusion detection technique. In the beginning, the high dimensionality problem can be addressed by the design of the CHHOFS method, which elects an optimum subset of features. Next, the attention‐driven ConvLSTM Autoencoder (ACLSTM-AE) model is employed for intrusion detection. Finally, the detection rate of the ACLSTM-AE method can be devised by the symbiotic organism search (SOS) model. The simulation results of the CHHOFS-ADL method have been validated on benchmark databases. A wide range of experiments highlighted the optimum performance of the CHHOFS-ADL method with recent models under various measures.