Paper Title: A deep learning-driven adaptive edge intelligence framework for real-time Internet of Things applications
Authors: Majd S. Ahmed
Corresponding Author: Majd S. Ahmed (majdsami@uomustansiriyah.edu.iq)/Iraq
Abstract
The rise of Internet of Things (IoT) ecosystems and the New Paradigm of low-latency edge data processing. IoT ecosystems are growing exponentially, and this growth brings a new challenge: low-latency data processing at the network edge. Because cloud-based architectures cannot meet this real-time requirement due to network delay, connectivity issues, and increasing bandwidth contention, deploying deep learning-based, high-computational-load pattern detection on high-dimensional streaming IoT data becomes challenging. The proposed framework adapts the model’s computational complexity and inference location in response to system state to optimize latency and energy while retaining high predictive accuracy. Experimental evaluation on the NSL-KDD intrusion detection dataset shows high accuracy, with class-wise accuracy of 0.979, precision of 0.981, recall of 0.977, F1-score of 0.979, and a precise ROC area under the curve of 0.998. Comparisons with simulation results provide useful insights into the relative advantages of local edge inference over end-to-end cloud inference, yielding 98% energy savings and a 97.6% latency reduction. This work reports a novel context-aware integrated framework that combines a lightweight deep learning classifier, adaptive task offloading using a Deep Q-Learning-based intelligent agent, hierarchical workload distribution at edge-fog-cloud levels, and proactive contextual information dissemination in real time, in order to ensure a context-aware system with 58.5%, 29.0%, and 12.5% edge, fog, and cloud-based decisions, respectively.
