The rapid spread of fake news on social media has caused great harm to society in recent years, which raises the detection of fake news as an urgent task. Recent methods utilize the interactions among different entities such as authors, subjects, and news articles to model news propagation as a static heterogeneous information network (HIN). However, this is suboptimal since fake news emerges dynamically, and the latent chronological interactions between news in HIN are essential signals for fake news detection. To this end, we model the dynamics of news and associated entities as a News-Driven Dynamic Heterogeneous Information Network (News-DyHIN), where the temporal relationships among news articles are well captured with meta-path based temporal neighbors. With the support of News-DyHIN, we propose a novel fake news detection framework, named D ynam i c H ierarchical A ttention N etwork (DiHAN), which learns news representations via a hierarchical attention mechanism to fuse temporal interactions among news articles. In particular, DiHAN first employs a temporal node level attention to learn the temporal information from meta-path based news neighbors through the modeled News-DyHIN. Then, a semantic attention layer is adopted to fuse different types of meta-path based temporal information for news representation learning. Extensive evaluations conducted on two public real-world datasets demonstrate that our proposed DiHAN achieves significant improvements over established baseline models.

The robustness of large language models (LLMs) becomes increasingly important as their use rapidly grows in a wide range of domains.Retrieval-Augmented Generation (RAG) is considered as a means to improve the trustworthiness of text generation from LLMs.However, how the outputs from RAG-based LLMs are affected by slightly different inputs is not well studied.In this work, we find that the insertion of even a short prefix to the prompt leads to the generation of outputs far away from factually correct answers.We systematically evaluate the effect of such prefixes on RAG by introducing a novel optimization technique called Gradient Guided Prompt Perturbation (GGPP).GGPP achieves a high success rate in steering outputs of RAG-based LLMs to targeted wrong answers.It can also cope with instructions in the prompts requesting to ignore irrelevant context.We also exploit LLMs' neuron activation difference between prompts with and without GGPP perturbations to give a method that improves the robustness of RAG-based LLMs through a highly effective detector trained on neuron activation triggered by GGPP generated prompts.Our evaluation on open-sourced LLMs demonstrates the effectiveness of our methods.