With the rapid development of generative adversarial networks, many image restoration problems that are difficult to solve based on traditional methods have gained new research approaches. With its powerful generation ability, generative adversarial networks can restore intact images from damaged images, so they are widely used in image restoration. In order to summarize the relevant theories and research on the problem of using generative adversarial networks to repair damaged images in recent years, based on the categories of damaged images and their adapted repair methods, the applications of image restoration are divided into three main aspects: image inpainting, image deblurring, and image denoising. For each aspect, the applications are further subdivided through technical principles, application objects and other dimensions. For the field of image inpainting, different image completion methods based on generative adversarial networks are discussed from the perspectives of using conditional guidance and latent coding. For the field of image deblurring, the essential differences between motion blurred images and static blurred images and their repair methods are explained. For the field of image denoising, personalized denoising methods for different categories of images are summarized. For each type of applications, the characteristics of the specific GAN models employed are summarized. Finally, the advantages and disadvantages of GAN applied to image restoration are summarized, and the future application scenarios are prospected.

Large language models (LLMs) have demonstrated significant potential and effectiveness across multiple domains. To evaluate the performance of mainstream LLMs in public security tasks, this study aims to construct a specialized evaluation benchmark tailored to the Chinese public security domain—CPSDbench. CPSDbench integrates datasets related to public security collected from real-world scenarios, supporting a comprehensive evaluation of LLMs across four key dimensions: text classification, information extraction, question answering, and text generation. Furthermore, this study introduces a set of innovative evaluation metrics designed to more precisely quantify the efficacy of LLMs in executing tasks related to public security. Through the in-depth analysis and evaluation conducted in this research, we not only enhance our understanding of the performance strengths and limitations of existing models in addressing public security issues but also provide references for the future development of more accurate and customized LLM models targeted at applications in this field.

现有恶意短信检测方法往往聚焦于提升检测准确率或速度,而忽略了模型自身的安全问题,因此,在真实场景中可能会遭受对抗样本攻击.为了解决上述问题,文章提出了一种融合对抗增强和多任务优化的恶意短信检测模型.在输入阶段,利用随机匹配池生成"原始文本-对抗样本"对作为输入,并且引入语义类型编码技术帮助模型区分数据边界.然后,使用基于ChineseBERT的单塔神经网络作为主干模型对短信的语义、拼音和字形特征进行挖掘.在输出阶段,使用监督的分类交叉熵损失和无监督的输入一致性损失作为多任务优化目标,以帮助模型学习文本对内在的关联特征并完成分类.基于公开数据集的实验结果表明,该方法的准确率和鲁棒性优于多种机器学习和深度学习检测方法.

池化算法是卷积神经网络中用于特征降维、参数压缩、扩大感受野的重要一层.针对现有的池化方法没有充分考虑到池化前特征图的整体内容及风格特征分布问题,提出了一种通过Sobel算子对卷积后的特征图计算每个特征点的梯度值,并根据梯度值分布确定每个池化窗口取最大值、均值或者最小值的池化算法.该算法充分考虑了特征图池化前后的整体内容及风格特征分布,保持了特征图的整体不变性.实验表明,该池化算法在VGG、ResNet等经典网络架构上取得了优异性能,具有普适性,可用来替代常用的最大池化、平均池化.

Pooling is an important part of modern convolutional neural networks, which can expand the perception field, reduce the parameter matrix, and avoid overfitting. Currently used maximum pooling, average pooling and various subsequent improved pooling algorithms cannot take into account the contour and background information of the feature map at the same time. In addition, the performance of different pooling algorithms varies greatly on different models and datasets. In this paper, we propose a learnable pooling algorithm. The introduction of learnable parameters allows the pooling layer to adaptively optimize the selection of key feature information that is beneficial to improve model performance during model training. It is experimentally verified that the pooling algorithm has superior performance over the existing maximum pooling and average pooling on several classical models and public datasets for image classification and text classification. The pooling algorithm with the introduction of learnable parameters can better prevent overfitting, steadily improve the accuracy of the model, and is highly generalizable.

Network embedding has facilitated lots of network analytical tasks by representing nodes as low-dimensional vectors. As an extension of convolutional neural networks from Euclidean data to irregular data, Graph Convolutional Networks (GCNs) provide a novel way to learn network representations and have attracted widespread attention currently. Most of the existing GCNs are only applied to unsigned networks. However, networks could have both positive and negative links in the real world, to which the unsigned algorithms are no longer applicable. In this paper, we propose a novel Signed Directed Attention Network model to capture the structural and social theoretical information of signed directed networks comprehensively through an auto-encoder framework. In the encoding block, the information of sign, direction, social theory, and "bridge" edges are encoded into node embeddings by a node fine-grained classification aggregation layer. Besides, a direction parameterization layer is also introduced to convert directions into direction-specific convolutional filters to enhance the node embeddings. In the decoding block, loss functions are designed to model sign, direction, "bridge" edge, and social theory information accordingly and make them complementary to each other to capture the network information fully. Experimental results for the signed link prediction task on several real-world signed directed graphs show that the proposed framework can achieve state-of-the-art performance.

Federated learning is a methodology that solves the contradiction of big data between user privacy and data sharing,and realize the concept of “data is invisible but available”.However,the federated model is at risk of backdoor attacks in the training process.The attacker trains a attack model containing a backdoor task locally,and amplifies the model parameters by a certain proportion to implant the backdoor into the federated model.Facing the backdoor threat in the training process of the horizontal federated learning,from the perspective of the game theory,this paper proposes a backdoor defense strategy and technical proposal based on the combination of random cutting and gradient clipping.After receiving the gradient from the participants,the central server randomly chooses the neural network layer from each participant,and aggregates the gradient contributions of each participant layer by layer.Then,the central sever clips gradient parameters according to gradient threshold.Gradient clipping and random cutting can weaken the influence generated by abnormal data from minority participants.It falls into platform state when the federated model learning backdoor features,so that it keeps failing on learning backdoor features without affecting the lear-ning process of target tasks.If the central server completes the federated learning during platform state,it can defend against backdoor attacks.Experimental results show that the proposed method can effectively defend against potential backdoor threats in fe-derated learning.At the same time,the accuracy of the model is ensured.Therefore,it can be applied in horizontal federated lear-ning scenarios,providing security protection for federated learning.

Natural language processing models based on deep learning have been widely used to analyze Chinese social media texts, such as Sina Weibo sentiment analysis, rumor detection, and news topic classification. However, while these analysis models have made breakthroughs in performance, there are also potential adversarial attack risks. To test the robustness of these models, an adversarial examples generation method CSMTP is proposed in this paper. The algorithm includes two main parts: a key to-kens searcher and an adversarial noise selector. The key tokens searcher is used to locate the key tokens that significantly affect the classification results when the internal details of the model are unknown, and the visual perception weight is introduced to measure human attention to these keywords. The adversarial noise selector provides a variety of adversarial noise generation strategies. It uses the policy network based on deep reinforcement learning training to select the appropriate type of noise according to the features of key tokens and target sentences. Experiments based on four public datasets show that CSMTP can launch effective adversarial attacks on Chinese social media text classification models, which has specific en-lightening significance for the follow-up research on the reliability of these models.

以Android 11和小米智能路由器为例,提出了在持久随机MAC地址背景下认定Android手机曾与目标AP连接的具体方法,为公安机关办理相关案件提供了参考.首先,详细剖析了Wi-Fi系统和Android存储系统;其次,研究了Android一般取证流程;最后,详细探讨了Android系统Wi-Fi取证和小米路由器取证的具体方法,并提出了在持久随机MAC地址背景下认定Android手机曾与目标AP连接的具体方法.

鉴于Pix2Pix在图像风格转换等图像翻译任务中存在细节丢失、生成图像模糊等问题,无法满足当前人脸素描生成任务的目标要求,提出一种改进的Pix2Pix模型.通过引入基于自注意力机制的残差卷积模块,让Pix2Pix的生成器和鉴别器在训练过程中能够为人脸图像的不同区域和通道赋予不同的权重,从而提高生成的人脸素描图像的质量,并且对Pix2Pix生成器的损失函数进行改进,使其生成的人脸素描图像更具有手绘风格.同时,针对生成对抗网络训练困难的问题,对原Pix2Pix的训练方法进行了改进.通过与Pix2Pix和CycleGAN对比,使用改进的Pix2Pix模型在训练过程中损失函数收敛更快、收敛过程更稳定,且生成的人脸素描图像在细节保留、轮廓清晰度等方面优于原Pix2Pix等模型,验证了改进Pix2Pix模型在人脸素描生成任务中的有效性.

The importance of power plant safety is increasing in the era of gradual technological development. When a fire occurs in the power plant, it will cause huge material losses, social unrest, and even casualties. The paper studies the common methods and models of fire warning, and introduces several model recognition techniques based on flames or smoke. Improved an automated power plant identification system based on the vision transformer, and proved the advantages of the technology through comparative analysis.

Face editing is widely used in public security pursuits,face beautification and other fields.Traditional statistical me-thods and prototype-based methods are the main means to solve face editing.However,these traditional technologies face pro-blems such as difficult operation and high computational cost.In recent years,with the development of deep learning,especially the emergence of generative networks,a brand new idea has been provided for face editing.Face editing technology using deep generative models has the advantages of fast speed and strong model generalization ability.In order to summarize and review the related theories and research on the use of deep generative models to solve the problem of face editing in recent years,firstly,we introduce the network framework and principles adopted by the face editing technology based on deep generative models.Then,the methods used in this technology are described in detail,and we summarize it into three aspects:image translation,introduction of conditional information within the network,and manipulation of potential space.Finally,we summarize the challenges faced by this technology,which consists of identity consistency,attribute decoupling,and attribute editing accuracy,and point out the issues of the technology that need to be resolved urgently in future.

深度学习模型被证明存在脆弱性并容易遭到对抗样本的攻击,但目前对于对抗样本的研究主要集中在计算机视觉领域而忽略了自然语言处理模型的安全问题.针对自然语言处理领域同样面临对抗样本的风险,在阐明对抗样本相关概念的基础上,文中首先对基于深度学习的自然语言处理模型的复杂结构、难以探知的训练过程和朴素的基本原理等脆弱性成因进行分析,进一步阐述了文本对抗样本的特点、分类和评价指标,并对该领域对抗技术涉及到的典型任务和数据集进行了阐述;然后按照扰动级别对主流的字、词、句和多级扰动组合的文本对抗样本生成技术进行了梳理,并对相关防御方法进行了归纳总结;最后对目前自然语言处理对抗样本领域攻防双方存在的痛点问题进行了进一步的讨论和展望.

针对发现和阻断COVID-19在人群中的传播问题,研究分析了谷歌和苹果公司在EN(Exposure Notification)项目中所提出的"去中心化"技术,详细探讨项目中的加密流程和密切接触者检测过程;对比分析了EN服务技术与我国健康码技术方案有效性、精准性和隐私性的强度,提出一种基于EN技术的新型健康码多维数据碰撞模型,旨在进一步提高疫情防控的强度和精准度,为完善我国疫情防控体系提供参考.

Hierarchical structures are very common in knowledge graphs, and semantic hierarchy-preserved knowledge graph embeddings have achieved promising results in the knowledge graph link prediction task. However, handling one-to-many, many-to-one, and many-to-many relations that can provide hierarchical information is challenging and brings entity indistinguishability issues. To address this limitation, this paper proposes a novel knowledge graph embedding model, namely Entity-context and Relation-context combined Knowledge Graph Embeddings (ERKE), in which each relation is defined as a rotation with variable moduli from the source entity to the target entity in the polar coordinate system. It can be seen as a combination of two spaces-modulus space and phase space. In the modulus space, modulus information is used to model semantic hierarchies, and entity-context information is adopted to make node representations more expressive. Besides, based on the design of the propagation rule of Graph Convolution Network (GCN), a new GCN model suitable for processing semantic hierarchies in knowledge graphs is proposed. In the phase space, relation-context information is used to make entities easier to distinguish. Specifically, a rotation operation in the polar coordinate system is transformed to the addition operation in the rectangular coordinate system, and relations between entities are mapped into their entity-specific hyperplanes. The proposed method is verified by the experiments on three benchmark datasets, and experimental results demonstrate that the proposed method can learn the semantic hierarchies in knowledge graphs and improve the prediction accuracy of complex one-to-many, many-to-one, and many-to-many cases simultaneously.

The numerical transformation of text is a key step in natural language processing tasks, among which the word embedding model is the most representative. However, the word embedding model is insufficient in representing unregistered and low-frequency words, and the character-level embedding model makes up for it. Most Chinese character-level models focus on the independent use of Chinese character features such as strokes, radicals, and pinyin, or the shallow correlations between some features, while the inherent correlations among different features such as pronunciation, glyph, stroke order, and word frequency are not fully utilized. Through the statistical analyses of various features of Chinese characters, this paper proposes a precoding method based on Character Helix Alternative Representation Model (CHARM), which can realize the reversible mapping of Chinese characters or words to English-like sequences, and the advantage of this method is verified in three tasks: text classification, named entity recognition and machine translation. Experimental results on several test sets show that the model performs well, and can be a replacement character-level corpus for the original Chinese text.

为解决汉语自然语言处理任务中未登录词问题,人们经常利用汉字的笔画、偏旁、拼音等细粒度特征提高模型的学习能力.为找出这类特征的最佳组合,通过统计方法研究了汉字的音节、起笔、偏旁、声调、词频、笔画数等特征,提出一种可融合多种汉字特征的跨象限助记符映射模型,即相码模型,该模型可自动实现中文字、词与字母编码间的可逆映射.在字符级模型的文本分类实验中,效果理想.此外,模型生成的编码长度适中,保留了可读特性,可用于特殊场合的文本标注,也能为中文文本提供等量的平行语料数据.可见,相码模型是自然语言处理中一个较好的辅助模型.

针对经典循环卷积神经网络(RCNN)在池化层采用的最大池化策略较为单一,会忽略除最突出特征外的其他特征,影响分类精度的问题,提出基于多头注意力池化的循环卷积神经网络(MHAP-RCNN)模型.多头注意力池化可以充分考虑各特征对分类的贡献,且能在训练过程中动态优化,有效缓解最大池化的单一性问题.在三个公开的文本分类数据集上进行实验,结果表明与经典RCNN及其他各模型相比,提出的模型具有更好的文本分类性能.

针对现有的网络病毒传播模型不适应自组织网络的问题,以无线传感器网络作为自组织网络的背景,提出一种复杂环境下的自组织网络病毒传播模型.该模型引入描述节点通信窗口开放情况参数以模拟休眠机制;引入描述网络链路好坏的参数以模拟环境影响因素;针对传感器节点资源有限等特征,提出一种抑制病毒传播的方案;使用在网络节点中注入补丁包的方式,完成无线传感器网络对病毒的免疫.通过与经典模型的对比实验证实了该模型对自组织网络具有良好的适应性.

Graph convolutional networks (GCNs) have been developed as a general and powerful tool to handle various tasks related to graph data. However, current methods mainly consider homogeneous networks and ignore the rich semantics and multiple types of objects that are common in heterogeneous information networks (HINs). In this paper, we present a Heterogeneous Hyperedge Convolutional Network (HHCN), a novel graph convolutional network architecture that operates on HINs. Specifically, we extract the rich semantics by different metastructures and adopt hyperedge to model the interactions among metastructure-based neighbors. Due to the powerful information extraction capabilities of metastructure and hyperedge, HHCN has the flexibility to model the complex relationships in HINs by setting different combinations of metastructures and hyperedges. Moreover, a metastructure attention layer is also designed to allow each node to select the metastructures based on their importance and provide potential interpretability for graph analysis. As a result, HHCN can encode node features, metastructure-based semantics and hyperedge information simultaneously by aggregating features from metastructure-based neighbors in a hierarchical manner. We evaluate HHCN by applying it to the semi-supervised node classification task. Experimental results show that HHCN outperforms state-of-the-art graph embedding models and recently proposed graph convolutional network models.