مرکز منطقه ای اطلاع رسانی علوم و فناوری Journal of Information Systems and Telecommunication (JIST) 2322-1437 14 53 2026 6 29 Robust Hybrid Deep Learning for IoT Unknown Intrusion Detection Under Data Scarcity Robust Hybrid Deep Learning for IoT Unknown Intrusion Detection Under Data Scarcity 1 15 en Ali Maroosi University of Torbat Heydarieh Amir Hossein Hojatinia University of Torbat Heydarieh Arash Deldari University of Torbat Heydarieh 2025 8 5 <p>The rapid expansion of the Internet of Things (IoT) has significantly heightened the need for robust intrusion detection systems capable of identifying previously unseen cyber threats. Traditional approaches often struggle with novel attack patterns, leading to decreased detection rates and increased system vulnerability. To address this critical limitation, we propose an innovative and highly effective framework that combines multi-source transfer learning with autoencoders to detect unlabeled and unknown attack types with exceptional precision. Unlike prior methods that rely on single-source transfer learning or basic feature fusion techniques, our advanced approach introduces two groundbreaking techniques: the Concurrent Feature Fusion Model (CoFFM) and the Cascading Feature Fusion Model (CaFFM). These models, along with an enhanced Unified Feature Fusion Model (UFFM), leverage autoencoders to significantly improve adaptability across diverse feature domains, ensuring superior performance in dynamic threat environments. Experimental results on benchmark datasets demonstrate that CoFFM achieves an outstanding accuracy rate of 98.13%, surpassing both non-transfer learning methods (92%) and the best single-source transfer learning approaches (94%). Furthermore, CoFFM exhibits remarkable efficiency under challenging conditions, achieving a substantial 12.24% performance gain over baseline methods even when trained on only 10% of the available data through random sampling. This highlights the model's exceptional robustness in data-scarce scenarios, making it a highly reliable solution for real-world IoT security applications. The success of our framework underscores the potential of multi-source transfer learning combined with autoencoder-based feature fusion in advancing the field of intrusion detection.</p> http://jist.ir/en/Article/Download/51082