How does IoT-XPLAIN ensure the reliability and faithfulness of GPT-4o–generated explanations to the underlying Random Forest decisions, especially under noisy inputs or adversarially injected faults, and how is hallucination or explanation drift mitigated?
IoT-XPLAIN enforces explanation faithfulness through strict architectural grounding and constrained prompting. The LLM does not access raw logs independently; it receives only structured inputs derived from the Random Forest model, including the predicted class, confidence score, top-ranked feature importances, and the normalized feature vector. Prompts explicitly instruct the model to base explanations solely on these provided signals, preventing external inference. Low-temperature decoding reduces variability, and explanations are generated only after a confirmed anomaly classification, avoiding speculative output. Under noisy or adversarially injected faults, robustness is preserved at the detection layer through SMOTE-balanced training and feature selection, while explanation drift is mitigated by limiting the LLM to summarization and structured reasoning rather than open-ended analysis. This tight coupling between model evidence and explanation output ensures consistency and minimizes hallucination risk.
