A Generalized, Lightweight, and Explainable Deep Learning Framework for Intrusion Detection Using Transfer Learning and CNN–BiLSTM Architecture

Abstract Intrusion Detection Systems (IDS) play a pivotal role in assisting modern networks, but their effectiveness is being compromised due to heterogeneous traffic conditions evolving cyber-attacks and incessantly growing deployment requirements in low-resource environments. The research closely unveils a deep learning-based IDS layout; making it general, adaptive, and explainable to keep high-detection accuracy under diverse or continuously altering network settings. The first stage refers to maintaining the appropriate generalization level for the model using transfer learning and adaptive feature representation. It is meant to enhance their ability to tolerate and work correctly against new or transforming types of attacks. The second stage of work sees the development of a light-weighted IDS architecture, which squares to operate in real time, to be executed within low-power platforms such as those of IoT and edge devices. The model will capture spatial characteristics with convolutional layers and temporal attack behaviors using bi-directional recurrent layers. To accommodate the lowering computational overhead, numerous model optimization strategies pruning, quantization, dimensionality reduction, feature selection either through Genetic Algorithm (GA) or Recursive Feature Elimination (RFE)—are applied. Deployment simulations mean to measure the suitableness of the model to real-time requirements. The metrics would cover latency, throughput, and energy consumption of the IDS model. In the last phase of the research, models are deployed with Explainable AI (XAI) technologies; more specifically, LIME and SHAP for improving interpretability of decisions and decision-making transparency. Here, many feature attributions are visualized with intrusion data. Interpretability gauged through fidelity, comprehensibility, and expert trust metrics. As such, these atlases can prove a unique example of how human understanding can be leveraged in support of decision-making within cybersecurity landscape. Keywords Adaptive IDS, CNN–BiLSTM, Transfer Learning, Lightweight Deep Learning, Feature Optimization, Real-Time Detection, Explainable AI (XAI)