SPARTA-net: Multi-modal transformer framework with cognitive vulnerability indexing for real-time social engineering attack detection in healthcare IoT and neural rehabilitation systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The proliferation of Internet of Medical Things (IoMT) devices, brain-computer interfaces (BCIs), and tele-rehabilitation platforms has introduced critical cybersecurity vulnerabilities, particularly social engineering attacks that exploit the cognitive limitations of neurologically impaired users. Existing intrusion detection and anomaly detection systems rely on static, rule-based paradigms that cannot model the dynamic socio-psychological manipulation strategies employed by advanced threat actors in healthcare IoT environments. This paper proposes SPARTA-Net (Socio-Psychological AI for Real-Time Threat Assessment Network), a novel deep learning framework that integrates a Multi-Head Cross-Attention Deception Detector (MHCADD) with a real-time Cognitive Vulnerability Index (CVI) for adaptive social engineering attack detection across textual, prosodic, and behavioral biometric modalities. The MHCADD employs cross-attention transformer architecture to capture inter-modal deception cues, while the CVI quantifies user vulnerability using keystroke dynamics, mouse trajectory irregularity, and interaction timing entropy. A gated contextual risk fusion engine combines these signals with environmental features to generate adaptive, user-specific threat scores. A continual learning module based on knowledge distillation addresses concept drift in evolving attack strategies. Experimental evaluation on the publicly available Phishing Email Dataset (available at: https://www.kaggle.com/datasets/naserabdullahalam/phishing-email-dataset ), comprising annotated phishing and legitimate email communications drawn from the Enron, Nazario, CEAS, Ling, and SpamAssassin corpora, with behavioral biometric streams augmented from the CMU Keystroke Dynamics Benchmark Dataset, demonstrates that SPARTA-Net achieves 96.73% detection accuracy, an 8.4% improvement in F1-score, and a 12.7% reduction in false positive rate compared to state-of-the-art baselines including BERT-Phish, DeepSE-Guard, and SE-Transformer. Real-time inference latency of 23.4 ms on GPU hardware suggests potential deployment feasibility in edge-constrained IoMT environments, subject to further clinical validation. Statistical significance (p 1.8) validates the superiority of the proposed framework. SPARTA-Net represents an offline computational proof of concept evaluated entirely on the publicly available SEA-Dataset benchmark; all reported performance metrics are obtained under controlled experimental conditions, and clinical viability remains subject to prospective validation in live tele-rehabilitation environments with certified IoMT hardware.