A Multi-Frequency Temporal Spatio-Transformer for adversarially robust IoT intrusion detection

The increasing prevalence of adversarial evasion techniques poses a significant challenge to the reliability of intrusion detection systems (IDSs) in Internet-of-Things (IoT) environments. Minor manipulations of traffic telemetry can alter temporal behaviour or frequency-level patterns, leading to degraded detection performance and potential security failures. This study presents the Multi-Frequency Temporal Spatio-Transformer (MFTST), a computational framework designed to improve adversarial robustness in IoT intrusion detection. MFTST combines temporal Transformer-based encoding with multi-frequency channel attention to capture both sequential traffic behaviour and frequency-level variations in learned traffic representations. The framework is evaluated using a Composite attack that emulates the detection-side effects of two representative defense-evasion behaviours: obfuscation (T1027) and indicator tampering (T1070), as commonly described in the MITRE ATT&CK knowledge base. The Composite attack jointly perturbs temporal attention behaviour and frequency-domain traffic representations to assess the resilience of MFTST against timing-based masking, attention disruption, and frequency-domain perturbations. To improve robustness, the model is trained using a composite adversarial optimization strategy that jointly considers classification performance, temporal-attention stability, and frequency-domain perturbation control. Experimental evaluations on MQTTset and X-IIoTID demonstrate that MFTST improves detection performance under both clean and adversarial conditions. The results show that adversarial training improves multiclass accuracy from 0.8314 to 0.8558 on MQTTset and from 0.8758 to 0.9192 on X-IIoTID. Under stronger adversarial evaluation, MFTST also maintains higher robustness than conventional machine-learning, deep-learning, and robustness-oriented baselines. These findings indicate that temporally conditioned frequency attention can improve both detection reliability and interpretability in adversarial IoT security monitoring.