The rapid development of information technology has made cyberspace security an increasingly critical issue. Network intrusion detection methods are practical approaches to protecting network systems from cyber attacks. However, cyberspace security threats have topological dependencies and fine-grained attack semantics. Existing graph-based approaches either underestimate edge-level features or fail to balance detection accuracy with adversarial robustness. To handle these problems, we propose a novel graph neural network–based method for network intrusion detection called the adversarial hierarchical-aware edge attention learning method (AH-EAT). It leverages the natural graph structure of computer networks to achieve robust, multi-grained intrusion detection. Specifically, AH-EAT includes three main modules: an edge-based graph attention embedding module, a hierarchical multi-grained detection module, and an adversarial training module. In the first module, we apply graph attention networks to aggregate node and edge features according to their importance. This effectively captures the network’s key topological information. In the second module, we first perform coarse-grained detection to distinguish malicious flows from benign ones, and then perform fine-grained classification to identify specific attack types. In the third module, we use projected gradient descent to generate adversarial perturbations on network flow features during training, enhancing the model’s robustness to evasion attacks. Experimental results on four benchmark intrusion detection datasets show that AH-EAT achieves 90.73% average coarse-grained accuracy and 1.45% ASR on CIC-IDS2018 under adversarial attacks, outperforming state-of-the-art methods in both detection accuracy and robustness.

Large Language Models (LLMs) have been proven remarkable in natural language processing, which prompts numerous works on knowledge extraction, knowledge fusion, knowledge representation, and knowledge completion. Existing works mainly focus on static multi-relational knowledge graph (KG). Unlike static knowledge graph, temporal knowledge graph (TKG) contains temporal information and evolves over time. Learning and reasoning about the representation of temporal knowledge graph are more difficult. Training or fine-tuning LLMs for temporal graph related tasks incurs significant computational overhead and requires the design of prompts. Conducting tasks of TKG does not necessarily require such complex work. Therefore, we explore temporal knowledge graph completion (TKGC) based on pre-trained "small" language model. We propose TKG-BERT by applying BERT for temporal knowledge graph completion and classification. Specifically, We introduce three ways to model temporal knowledge in TKG-BERT: vanilla knowledge embedding (Van.), explicit time modeling (Exp.) and implicit time modeling (Imp.). TKG-BERT(Van.) only adopts static knowledge without embedding time information; TKG-BERT(Exp.) embeds timestamp in quadruple explicitly; TKG-BERT(Imp.) models time implicitly, by dividing the training set and testing set in chronological order. We conduct experiments on ICEWS14 and ICEWS05-15, which are two public temporal knowledge graph datasets. Various experiments of temporal knowledge graph completion and classification tasks show the effectiveness of pre-trained language model for TKG completion. We also compare the performance of TKG-BERT accross different time modeling way and proportion of training set.

Temporal Knowledge Graph (TKG) reasoning aims to infer time-dependent facts by modeling how entities and relations evolve. Most existing methods learn entity and relation embeddings jointly from past observations and often treat relations as auxiliary signals induced by entity interactions. This coupling can over-smooth relation representations and weaken the model’s ability to capture relation-specific semantics, logical dependencies, and context shifts that are crucial for predicting future facts. To overcome this limitation, we propose TECA (Temporal Evolution with Convolutional Block Attention), a relation-centric module designed to explicitly model relation evolution in temporal graphs. TECA builds time-aware relation representations by integrating static semantics with temporal dynamics, and employs a lightweight Convolutional Block Attention Module (CBAM) that combines channel attention and spatial attention. Channel attention adaptively reweights semantic feature dimensions, while spatial attention highlights informative local interactions, enabling the model to emphasize the most discriminative relational cues under changing temporal contexts. We evaluate TECA on four widely used TKG benchmarks (ICEWS14, ICEWS18, ICEWS05–15, and GDELT) covering different temporal granularities and relational complexities. Extensive experiments and component analyses show that TECA consistently surpasses recent state-of-the-art baselines, achieving improvements of up to 6.1 MRR points and 6.3 Hits@10 points. The results indicate that explicitly enhancing relation evolution is a key factor for robust and accurate TKG reasoning.