Hybrid deep-learning Network Intrusion Detection Systems (NIDS) achieve strong clean-data accuracy but typically leave adversarial robustness, deployment cost, and quantitatively verified explainability unaddressed. We formulate NIDS design as a constrained multi-objective learning problem spanning detection accuracy, adversarial robustness, latency/resource-bounded deployability, and explanation fidelity, and we design a corresponding attention-augmented architecture with SHAP explainability, FGSM/PGD adversarial training, and knowledgedistilled edge deployment. To validate the overall pipeline before a full-scale implementation on real benchmark traffic, we build, from scratch in NumPy, a lightweight prototype (per-timestep feature projection, a recurrent temporal encoder, single-head self-attention pooling, and a softmax classifier, trained via manually-derived backpropagation-through-time) and evaluate it end-to-end on a controlled synthetic multi-class flow-sequence dataset. The prototype reaches 94.7% test accuracy (macro-F1 0.947, AUC 0.970) on held-out synthetic data, degrades from 94.7% to 80.3% accuracy under an undefended FGSM attack as the perturbation budget grows, and a 9.2x smaller variant matches its accuracy while cutting measured inference latency by 2.5x. These genuine, reproducible findings validate the pipeline and concretely motivate the adversarial-hardening and distillation components of the full design, whose evaluation on real, large-scale NIDS benchmarks is identified as the next phase of this research.

Network intrusion detection, deep learning, recurrent neural network, attention mechanism, adversarial robustness, explainable AI, knowledge distillation, edge deployment, synthetic-data validation.