AdaStack-IoT: Adaptive Depth Multi-Tier Meta-Learning for Dynamic Intrusion Detection in Resource-Constrained IoT Networks

Abstract

The proliferation of Internet of Things (IoT) devices introduces critical security vulnerabilities, yet standard intrusion detection systems (IDS) are often impractical in resource-constrained environments due to high computational overhead and limited adaptability to novel threats. To address this, we propose AdaStack-IoT, a novel adaptive depth multi-tier meta-learning framework designed for dynamic, resource-aware intrusion detection. Our methodology employs a three-tier architecture: Tier-1 utilizes heterogeneous base learners (Random Forest, Gradient Boosting, lightweight CNNs, and LSTMs); Tier-2 integrates their predictions via an attention-weighted meta-learner; and Tier-3 features an adaptive depth controller using Proximal Policy Optimization (PPO) to dynamically adjust computational depth based on traffic complexity and real-time resource availability. We validated AdaStack-IoT within a MATLAB simulation environment using the BOT-IoT and UNSW-NB15 datasets. Experimental results demonstrate superior performance against five state-of-the-art baselines, achieving a detection accuracy of 98.7% and an F1-score of 98.4%. Critically, the adaptive mechanism yielded a 42% reduction in energy consumption and a 35% improvement in inference latency, all while maintaining high detection efficacy across eight attack types, including DDoS, botnet, and zero-day exploits. These findings highlight AdaStack-IoT’s significant potential as an efficient and robust security solution for IoT networks with limited resources.