Background: Machine learning-based intrusion detection systems are being used increasingly to keep internet of things (IoT) environments safe, but the fact that they can be tweaked by adversaries limits their dependability, especially in resource-constrained, real-time deployments.

Objective: This study proposes and evaluates an adversarially robust intrusion detection framework, integrated into the OneM2M service layer, for real-time edge IoT security, where the goal is to keep communications safe, even when attackers act in crafty ways.

Methods: Within the OneM2M analytics engine, the framework combines statistical anomaly filtering, feature squeezing, adversarial training, confidence driven rejection and federated learning. It is tested across white-box, black-box, transfer learning style and even data poisoning situations using IoT security datasets, OneM2M traffic patterns and edge-centred deployment setups.

Results: This layered defence approach improves resilience against adversarial attacks while keeping detection performance, inference efficiency and resource use within defined limits. In the evaluation part, it shows operational feasibility too, even across varying traffic loads and different edge deployment configurations.

Conclusion: Integrating a multilayer adversarial-defence pipeline into standardized OneM2M middleware can help reinforce real-time IoT intrusion detection while not completely giving up edge-resource limits. It is more robust, workable in practice and less easy for attackers to slip through in the usual environment.

Keywords: Federated learning; M2M analytics engine; Feature squeezing; Evasion attacks; Embedded devices; Confidence-based rejection; Internet of things.