Physics Inspired Optimisation and Explainable AI Framework for Enhanced BHP Flooding Attack Classification in Optical Burst Switching Networks

Abstract

Optical Burst Switching (OBS) networks offer high bandwidth efficiency and low latency, making it an ideal choice for next generation high-speed photonic communications. However, the burst based transmission architecture is highly vulnerable to flooding attacks, which can severely degrade network performance. In this work, a hybrid approach for Burst Header Packet (BHP) flooding attack classification is proposed. The method combines direct Machine Learning (ML) on tabular data, tabular to image conversion using EfficientNet-b0 fine-tuning. Further, deep EfficientNet-b0 features are optimized using physics inspired Black Hole Optimisation with Adaptive Mutation (BHO-AM). Finally, the optimised features are classified using a Bayesian-optimized Support Vector Machine (SVM) classifier. Explainable AI (XAI) techniques, including Grad-Class Activation Map (CAM) and Occlusion Sensitivity, are employed to enhance interpretability and identify the most critical features influencing classification. Experimental results show that Efficient Net fine-tuning achieves 99.50 % accuracy, while the BHO-AM optimized SVM model attains 99.60 % accuracy with significantly reduced training time. This study introduces a novel tabular to image conversion framework for OBS attack data, enabling Deep Learning models to achieve high accuracy. Thus, combining XAI methods, with DL classification the proposed work achieves better performance and enhanced interpretability in OBS networks.