By synergizing external knowledge with generative capabilities, RAG is emerging as an important approach for enhancing the professionalism and reliability of Large Language Models (LLMs). However, due to the difficulty of fully ensuring the reliability of external knowledge sources, RAG systems are vulnerable to data poisoning attacks. Existing poisoning attacks are typically designed for a single specific query, exhibiting limited generalization capability. In addition, existing methods craft poisoned texts with low retrieval probability, leading to limited attack effectiveness. What is more, they also suffer from inefficiency in poisoned text construction. In this paper, we propose AutoTPA, a novel RAG poisoning attack, which targets all user queries containing the specific trigger. AutoTPA integrates dynamic query clustering, efficient and minimal candidate token set construction, as well as heuristic token evaluation algorithm, enabling automated and efficient construction of effective poisoned texts. All the attacker has to do is to determine the trigger to attack. Extensive experimental results demonstrate that AutoTPA achieves an attack success rate up to 95.5 % in less time while using only a token set sized at about 3.5 % of the original vocabulary. It significantly outperforms existing methods, showing clear advantages in both efficiency and effectiveness.

Substructure mining serves as a fundamental technique for uncovering key patterns and latent risk within complex networks.Existing methods focus on modeling of topological and semantic information, while neglecting the cross-layer semantic confusion and initial decisions unreliability resulting from intrinsic coupling between them.Faced with that, this study proposes a novel Graph Transformer model based on fact reflection, named RAGphormer, which enhances substructure mining through high-order semantic modeling and dynamic decision reflection.Specifically, we introduce the Sub2Token module, which constructs hierarchical semantic encoding sequences by integrating node-level, substructure-level, and multi-hop neighborhood features.It overcomes the limitations of cross-layer semantic sharing among multi-hop neighborhood nodes, and mitigates the noise caused by multi-hop semantic entanglement.Furthermore, a dynamic decision reflection mechanism is proposed, which constructs substructure fact and counterfactual feature patterns based on an adaptive KMeans clustering algorithm.A dynamic thresholding strategy is employed to identify potentially unreliable predictions, which are then revised based on the discrepancies between corresponding feature patterns.Extensive experiments on two private datasets (ChTech and EnTech) and one public dataset (Cora) demonstrate that RAGphormer outperforms baseline methods across multiple evaluation metrics, achieving an accuracy of up to 98.8%.

Deep neural network are one of the most powerful model for machine learning, which can learn the underlying patterns automatically from a large amount of data. So it can be extensively used in more and more Internet-of-Things (IoT) applications. However, the training of deep models is difficult, suffering from overfitting and gradient vanishing problem. Besides, the large amount of parameters and multiplication operations make it impractical for most deep learning models to directly execute on target hardware. In this paper, we propose a method of gradually pruning the weakly connected weights to improve the traditional stochastic gradient descent. And we adopt a reinforcement learning method called learning automata to find the weakly connected weights on account of its strong policy-making ability in stochastic and nonstationary environment. Our proposed method can learn a more effective and sparsely connected architecture during training from the initially fully connected neural networks. The experiments on MNIST show that our method have stronger power to defeat overfitting and can get better generalization performance on test set. Meanwhile, the thin and sparsely connected model we get can be more suitable for IoT applications.

Along with the fast development in information and communication technology (ICT), job skills required by ICT industries are also evolving very rapidly. It becomes difficult for ICT students to assess the gap between their skills and such evolving skills. Even though schools perform periodical curriculum evaluations, the time gap between the evaluations causes the curriculum to get out-of-date easily since it is unable to cope with the tremendous and quick changes occurring in the industry. We propose novel solutions by introducing some measures and visualization tools to reveal such skills’ gap. Using evolutionary-based data mining, the skillsets mastered by students were collected from their study reports, while the frequent skillsets required by the industry were mined out from job adverts; and based on these skillsets the skill coverage of the students was approximated. The proposed solutions were then tested on data obtained from an Indonesian higher education institution since Indonesia implements competence-based curriculum in its education system. Experimental works show that the proposed approaches not only reveal and visualize the gap, but also monitor the changes in the skills requirements, which also help the school’s administrator while updating the curriculum.