Open-set IoT intrusion detection must distinguish unfamiliar attacks from uncertain known traffic. OpenTriIDS implements a three-action policy: a random forest accepts high-confidence predictions, routes uncertain samples to a compact known-class prototype memory, and rejects routed samples beyond a standardized Euclidean distance boundary. At the representative S1 operating points on Edge-IIoTset and CIC-IoT-2023, the operational macro-F1 reaches 68.44% and 76.81%, with routing rates of 4.53% and 4.13%, respectively. The representative S1 analysis shows that the confidence gate defines the selected subset, while the three-prototype support rule improves macro-F1 and changes the balance between known-class retention and unknown rejection; the magnitude and direction depend on the dataset and operating point. The prototype-distance module records the matched label and distance for each routed sample. A complementary controlled protocol evaluates three seeds across multiple held-out-family settings, while a seven-method S1 comparison evaluates MSP, OpenMax, ORI open recognition, EFC, post hoc Energy OOD, single-layer Mahalanobis scoring, and OpenTriIDS at their documented operating points. OpenTriIDS achieves the highest macro-F1 among the evaluated methods in both datasets under this comparison. Together, the analyses characterize a selective open-set policy whose matched-gate macro-F1 advantage is accompanied by dataset- and holdout-dependent unknown-recall trade-offs, while distinguishing predictive performance from routing and computational cost.
With the development of network technology, there is a large amount of traffic data in the network at all times. In order to maintain network order and prevent attackers from damaging it, it is necessary to study methods for real-time monitoring of network traffic anomalies. The current machine learning based network traffic anomaly detection methods mostly rely on complex feature dimensionality reduction and feature selection techniques. The model relies on labeled data and is prone to getting stuck in local optima during use. A self supervised patch Transformer(SSPT) based network intrusion detection method is proposed. To solve the problem of wide distribution and strong discreteness of raw attack data, the data is first subjected to single hot encoding and normalization processing, and then segmented into subsets as inputs to the Transformer. Then, training and classification are achieved by sharing the same weights through independent channels. The paper conducted experiments on the CIC-Bell-DNS-EXF-2021 dataset, and the results showed that the proposed method achieved an accuracy of 99.34% for the aforementioned dataset, which is 2.80% higher than the same type of LSTM.
Malicious traffic detection in IoT environments faces dual challenges: limited labeled data and heterogeneous, complex traffic patterns. To address these limitations, we propose a malicious traffic detection framework, GADF-SRGA, which integrates Gram-angle-difference-field (GADF) imaging with meta-learning. The framework first encodes raw IoT traffic into images via GADF, preserving the spatiotemporal characteristics of malicious traffic. It then employs meta-learning on these encoded images to enable feature-space learning under scarce data. In the inner loop, Sample-Relation Guided Attention (SRGA) leverages class-label-guided supervision graphs to learn sample similarity, improving intra-class compactness and inter-class separability in the feature space. Comprehensive evaluations on public IoT intrusion datasets Malicious_TLS and ToN_IoT demonstrate the framework’s superiority and robustness, particularly under class-imbalanced conditions, over baseline methods.
Unmanned aerial vehicle (UAV) networks are exposed to security threats arising from open wireless access, dynamic connectivity, and resource-constrained network devices. This study develops a lightweight Transformer-based network intrusion detection architecture with potential application to UAV-network security. The model combines rotary position embedding (RoPE), a local aggregation attention unit (LAAU), self-attention, and a lightweight feedforward network (FFN). Grouped linear transformations first aggregate local traffic representations, after which self-attention models global dependencies. The resulting representation is then refined by the lightweight FFN. Experiments on the CICIDS2017 and UNSW-NB15 benchmark datasets compare the proposed architecture with deep neural network (DNN), convolutional neural network (CNN), long short-term memory (LSTM), and hybrid CNN–gated recurrent unit (CNN-GRU) baselines. Noise-robustness, hyperparameter, and ablation analyses are also conducted. The results demonstrate the effectiveness of the proposed architecture in the tested benchmark settings. The results provide benchmark evidence for the development of intrusion detection methods with potential application to UAV networks.