Hybrid Quantum–Classical Deep Reinforcement Learning with Novel Micro–Macro Adaptive Time Stepping and Bio-Inspired Optimization for IoT Threat Detection

The exponential growth of Internet of Things (IoT) deployments has expanded the cyberattack surface and exposed resource-constrained devices to polymorphic, zero-day, and stealthy threats. Classical and deep-learning intrusion detection systems (IDS) face four persistent limitations: static decision boundaries, poor temporal modelling, high computational overhead, and brittle adaptation to evolving attack patterns. This paper proposes MMATS-QRL, a hybrid quantum–classical deep reinforcement learning framework for adaptive IoT intrusion detection. The framework contributes three new algorithms. First, a Micro–Macro Adaptive Time Step (MMATS) controller, inspired by heterogeneous multiscale numerical analysis, dynamically alternates between fine and coarse temporal resolutions according to a behavioural volatility metric derived from the Q-gradient. Second, an Adaptive Lizard Algorithm (ALA) performs feature selection by combining Lévy-flight foraging exploration, chromatic-adaptation exploitation, and tail-autotomy pruning. Third, an Adaptive Blanket Body Cover Algorithm (ABBCA) tunes hyperparameters through a thermoregulation-inspired coverage, equilibrium, and refinement schedule. Selected features are encoded into a parameterised quantum circuit and compressed into a scalar vulnerability score that becomes the state of a Deep Q-Network (DQN) agent, transforming intrusion detection into a sequential decision process. Experiments on NSL-KDD, CICIDS2017, UNSW-NB15, BoT-IoT, and IoT-23 give an average detection accuracy of 98.70 percent and F1-score of 98.50 percent, with a 67 to 84 percent training-time reduction and a 69 to 76 percent memory reduction relative to CNN, LSTM, and hybrid deep learning baselines. Ablation confirms that each component contributes independently: quantum encoding adds 2.34 percentage points of accuracy and MMATS yields an 18.7 percent training speed-up at equal accuracy. The quantum component is evaluated using classical state-vector simulation, so these gains reflect a quantum-inspired feature representation rather than execution on physical quantum hardware. The compact state space and adaptive temporal resolution make the framework suitable for edge-class deployment in heterogeneous IoT networks.