An Intelligent Framework for Early Detection of AI-Induced Risks in Safety-Critical Systems

Abstract

IRED-AI is an intelligent framework designed for the early detection of AI-induced risks in safety-critical systems, including healthcare, autonomous driving, aviation, industrial control, and cybersecurity. Unlike traditional detection models that optimize single metrics, IRED-AI unifies continuous monitoring, hybrid statistical–AI anomaly detection, explainable reasoning, and retrieval-guided adaptation into a cohesive, SLA-compliant architecture. A multi-view feature engineering pipeline captures contextual, signal, and semantic health indicators to construct risk embeddings that enable real-time anomaly scoring and efficient case retrieval. The framework’s adaptive alert escalation process provides operators with actionable insights at informational, cautionary, critical, and emergency levels, while integrated explainable AI modules ensure transparency and regulatory compliance. Best results across ten benchmark datasets confirm its superiority: accuracy of 95.20%, AUROC up to 0.960 (aviation), AUPRC gains of +3.37%, latency reduced to 90–150 ms (−20.83%), false alarms lowered to 7.20% (−22.22%), throughput reaching 1500 q/s, Top-1 retrieval accuracy of 92.40% (Top-5 = 97.20%), robustness of 0.920, interpretability of 0.900, and SLA compliance of 98.00%. Statistical analysis confirms these improvements as significant (p < 0.050), demonstrating that IRED-AI’s enhancements are not incidental but systematically robust. The framework also incorporates a knowledge retention loop, enabling continuous adaptation through operator feedback and periodic recalibration, ensuring resilience against data drift, adversarial manipulation, and emerging risks. By combining technical excellence with practical trustworthiness, IRED-AI provides a deployable, domain-general solution that advances safety assurance in mission-critical environments where failures carry unacceptable human and economic consequences.