Short-video platforms now host vast multimodal ads whose deceptive visuals, speech and subtitles demand finer-grained, policy-driven moderation than community safety filters. We present BLM-Guard, a content-audit framework for commercial ads that fuses Chain-of-Thought reasoning with rule-based policy principles and a critic-guided reward. A rule-driven ICoT data-synthesis pipeline jump-starts training by generating structured scene descriptions, reasoning chains and labels, cutting annotation costs. Reinforcement learning then refines the model using a composite reward balancing causal coherence with policy adherence. A multitask architecture models intra-modal manipulations (e.g., exaggerated imagery) and cross-modal mismatches (e.g., subtitle–speech drift), boosting robustness. Experiments on real short-video ads show BLM-Guard surpasses strong baselines in accuracy, consistency and generalization.

Low-light image enhancement (LLIE) aims at restoring the color and brightness of underexposed images. Supervised methods suffer from high costs in collecting paired low-normal light images, while unsupervised approaches require intricate loss functions. To tackle these dual challenges, we propose the Trouble-Maker Learning (TML) strategy, which leverages images with normal light as training inputs. TML comprises two core components. Firstly, the Troublemaker Model (TM) generates pseudo low-light images from normal images, thereby alleviating the need for pairwise data and reducing associated costs. Secondly, the Predicting Model (PM) enhances the brightness of pseudo low-light images. Additionally, we integrate an Enhancing Model (EM) to further refine the visual quality of the PM's outputs. In LLIE tasks, it is crucial to capture global element correlations, as this allows for the extraction of more information pertaining to the same object. Convolutional Neural Networks (CNNs) and self-attention mechanisms are not well-suited to this task due to the local CNN operators, and high time complexity, respectively. To address these limitations, we propose Global Dynamic Convolution (GDC) with a time complexity of O(n). Essentially, GDC mimics the partial calculation process of self-attention to establish element-wise correlations. Building upon the GDC module, we develop the UGDC model. Finally, we explore the application of Data Fusion in the field of LLIE. Based on the Retinex theory, we conducted feature-level fusion using low-light images, illumination components and reflection components, which further enhance the performance of the LLIE system. Extensive quantitative and qualitative experiments demonstrate that UGDC, trained with TML and via data fusion, can achieve performance competitive with state-of-the-art approaches on public datasets. The source code of this paper is publicly available at https://github.com/Rainbowman0/TML_LLIE, facilitating reproducibility of the research findings.

Graph node feature aggregation is a very common basic operation in graph data query and graph neural networks. We design three noised aggregation methods to protect the node features of the graph and analyze the privacy of these methods from the perspective of differential privacy. NAGG1, our first method, adds Gaussian noise to the aggregation result, extending the classic Gaussian mechanism of differential privacy directly to the graph domain. Subsequently, we design NAGG2 which adds noise to the raw data of each sample and then performs the aggregation. This makes it possible to decouple the data owner and subsequent data users. We impose some restrictions on the original definition of differential privacy, and under these new assumptions, we demonstrate that the effects of Gaussian noise in NAGG2 can be achieved with any other type of noise that meets the requirements. We call this method NAGG3, which significantly enhances the flexibility of added noise in theory. For these methods, we instantiate them with mean aggregation and graph convolutional aggregation. We analyze their privacy protection using the framework of differential privacy and evaluate their utility through experiments.