MaID-Agent: A Hybrid Masqueraded Intrusion Detection System Enhanced by Agent

The deployment of MQTT/IoT and the emergence of fine-tuned large language models (LLMs) enable a new evasion opportunity for attackers: a masquerade attack that uses a language model to generate fake-normal (F) traffic which mimics the activities of a normal broker and thereby dilutes the malicious (M) fraction seen by IDSs. Our analysis shows that existing single-modality detectors fail at representation in the presence of such camouflage and that this problem cannot be addressed merely through re-training. We formulate the fake-normal threat model and introduce MaID-Agent, a two-tier solution where in Tier-1, a multi-view representation that combines three independent features, temporal (TranAD), payload semantic (FAISS), and character-morphological (CharSVD) using an Adaptive Weighting Network (AWN) trained under an F-aware classification rule that maps both F and M instances into the same positive category. In Tier-2, we call an agent on the thin slice where there is uncertainty, resulting in F/M disambiguation and per-sensor evidence and autonomous actions including {IGNORE, BLOCK_ALERT, UPDATE_DB}. MaID-Agent is validated against seven single-modality IDSs across five seeds with respect to malicious-to-camouflage ratios ranging between 8:8 and 15:1 on MQTTSet and DDMI. While the recall performance of all the baseline detectors monotonically drops with increasing malicious-to-camouflage ratios. TranAD's recall drops to 0.356 (MQTTSet) and 0.200 (DDMI) at ratio 15:1. MaID-Agent restores to 0.986 (MQTTSet) and 0.988 (DDMI). The recovery is due to morphology (MQTTSet) and payload semantic (DDMI) feature, a clear demonstration of the effectiveness of multimodality rather than ensembling. In the Tier-2, the agent fills the decision-to-action gap based on the verdict provided by Tier-1 at sub-100ms amortized per-packet costs. These findings highlight the importance of a combination of multimodal detection with an agent for building robust IDS systems capable of operating under new and emerging adversarial conditions, ultimately paving the way for more resilient IoT security frameworks in the face of evolving threats.