Performance evaluation of DNN–LSTM–based optical communication monitoring for joint failure diagnosis and physical-layer attack detection

Abstract Accurate optical communication monitoring is essential for distinguishing optical transmission failures from physical-layer attacks that produce similar signal degradation characteristics. This study proposes a class-specific validation–optimized DNN–LSTM (CSVO-DL) framework for joint optical failure diagnosis and physical-layer attack detection. The framework considers seven operating states, comprising normal operation, three optical failure conditions, and three physical-layer attacks. Physically constrained attack scenarios are constructed from optical measurements, with corresponding bit-error rates estimated using a quality-of-transmission model. Unlike conventional fixed-weight fusion, the proposed framework integrates DNN and LSTM predictions through class-specific weights optimized using validation data to improve multiclass discrimination. Experimental evaluation on a balanced dataset of 910,000 samples, including an independent test set of 210,000 samples, achieved 91.190 % accuracy, 91.600 % macro-precision, and 91.056 % macro-F1 score, outperforming the evaluated standalone machine-learning and deep-learning baselines. The framework achieved macro-recall values of 98.468 % for optical failures and 80.553 % for physical-layer attacks. Ablation analysis demonstrated an improvement in macro-F1 from 90.819 % with fixed-weight fusion to 91.056 %, while introducing only 0.019 s of additional fusion overhead for the complete test set. These findings highlight the potential of class-specific deep-learning fusion for integrated optical performance monitoring, failure diagnosis, and physical-layer security.