As the proliferation of Consumer Internet of Things (CIoT) and Industrial IoT (IIoT) devices intensifies, ensuring secure and interpretable edge deployment of Intrusion Detection Systems (IDS) has become a critical challenge. The rapid expansion of CIoT/IIoT networks at the edge has introduced complex cyber-attack surfaces, posing significant challenges to conventional machine learning and deep learning-based IDS. These challenges include adversarial vulnerabilities where attackers deliberately inject malicious samples to mislead security predictions, class imbalance, and opaque black-box decision processes. Motivated by these challenges, this research proposes a novel unified eXplainable Artificial Intelligence (XAI) enhanced adversarial resilient deep learning framework for transparent, robust, and resource-efficient edge deployment in CIoT/IIoT environments. Unlike existing methods that rely on oversampling techniques such as SMOTE, our approach leverages Conditional Generative Adversarial Networks (CGANs) to synthetically balance highly imbalanced intrusion classes without relying on oversampling heuristics. We further design an LSTM-based denoising autoencoder to perform nonlinear dimensionality reduction, significantly improving noise robustness and edge deployability. In order to strengthen adversarial resilience, we introduce a defense strategy via Automatic Projected Gradient Descent (Auto-PGD), Square Attack, Carlini-Wagner (CW), and DeepFool during training to enhance robustness against both white-box and black-box perturbations. We built a hybrid Multi-Head Self-Attention (MHSA) and Bidirectional Gated Recurrent Units (BiGRU) for rich sequential learning and contextual sensitivity. Beyond performance, the model integrates Shapley Additive explanations (SHAP) to deliver both global and local post-hoc feature attributions, enhancing interpretability and trust. We evaluated our framework on two recent representative datasets Edge-IIoTset (IIoT-specific) and CIC-IoT2023 (general IoT) and demonstrated detection accuracies exceeding 99% on both clean and adversarial samples, with a minimal memory overhead of less than 140 MB and real-time inference latency of 32.1 ms per sample on the Raspberry Pi 4 and Jetson Nano. SHAP visualizations validate the framework’s decision rationale, highlighting protocol-aware and behaviorally critical features as dominant predictors, supporting forensic analysis across diverse attack types. Compared against state-of-the-art adversarial and non-adversarial models, our framework achieves superior performance in adversarial robustness, interpretability, and edge efficiency, without sacrificing detection precision. This research presents a deployable, auditable, and secure IDS solution aligned with standards such as CCPA, GDPR, NISTIR 8259, and HIPAA, advancing the frontier of intelligent edge security in CIoT/IIoT ecosystems.
Saheed et al. (Fri,) studied this question.