Electrical Engineering and Computer Science Student Publications
Document Type
Conference Proceeding
Publication Title
2026 IEEE 27th International Conference on Information Reuse and Integration for Data Science (IRI)
Abstract
Deploying large language models (LLMs) as decision-making agents in safety-critical Internet of Things (IoT) systems introduces risks that purely behavioral guardrails, such as action whitelists, cannot fully address. This paper presents Epistemic Edge, a four-tier neuro-symbolic pipeline that augments LLM-driven actuation with subjective logic (SL) uncertainty quantification, temporal decay, and dual guardrails combining epistemic threshold checks with behavioral whitelists. We evaluate seven locally-deployed models: three PrismML Bonsai 1-bit models (1.7B, 4B, 8B), three 4-bit quantized models (Qwen3-8B, Llama 3.2-3B, Phi-3.5-mini), and the DeepSeek-R18B reasoning model, across eight ablation conditions and five IoT actuation scenarios (2,800 controlled trials). We further validate the pipeline on the BATADAL water distribution cyberattack benchmark (4,177 hours of real sensor telemetry, 5 attack events) across 422 hyperparameter configurations. The evaluation yields three findings. First, removing the epistemic layer causes threshold guardrail accuracy to collapse from 1.00 to exactly 0.40 across all seven models, showing that uncertainty-aware guardrails are a necessary complement to behavioral whitelists. Second, the BATADAL evaluation reaches AUROC 0.9004 (no trained weights), with a significant fusion effect (paired Wilcoxon p < 10^-4, Cohen's d=1.01). Third, different BATADAL events are optimally detected by different SL-derived signals, with no single signal winning on all five -- a structural finding with implications for SL as an anomaly-detection framework. All code, experiment scripts, and result data are released as open-source.
DOI
10.1109/IRI69576.2026.00017
Publication Date
2026
Recommended Citation
Syed, Muntaser and Silaghi, Marius, "Epistemic Edge: Subjective Logic Guardrails for LLM-Driven IoT Actuation" (2026). Electrical Engineering and Computer Science Student Publications. 104.
https://repository.fit.edu/ces_student/104