2021 Vol. 58 No. 7
2021, 58(7): 1351-1352.
DOI: 10.7544/issn1000-1239.2021.qy0701
2021, 58(7): 1353-1365.
DOI: 10.7544/issn1000-1239.2021.20200979
Abstract：
With the rapid increase in access to the internet and the subsequent growth in the population of social media users, the quality of information posted, disseminated, and consumed via these platforms is an issue of growing concern. A large fraction of the common public turn to social media platforms and, in general, the internet for news and even information regarding highly concerning issues such as COVID-19 symptoms and treatments. Given that the online information ecosystem is extremely noisy, fraught with misinformation and disinformation, and often contaminated by malicious agents spreading propaganda, identifying genuine and good quality information from disinformation is a challenging task for humans. In this regard, there is a significant amount of ongoing research in the directions of disinformation detection and mitigation. In this survey, we discuss the online disinformation problem, focusing on the recent ″infodemic″ in the wake of the coronavirus pandemic. We then proceed to discuss the inherent challenges in disinformation research, including data collection, early detection and effective mitigation, fact-checking based approaches, multi-modality approaches, and policy issues and fairness, and elaborate on the interdisciplinary approaches towards the detection and mitigation of disinformation, after a short overview of the various directions explored in computational detection and mitigation efforts.
With the rapid increase in access to the internet and the subsequent growth in the population of social media users, the quality of information posted, disseminated, and consumed via these platforms is an issue of growing concern. A large fraction of the common public turn to social media platforms and, in general, the internet for news and even information regarding highly concerning issues such as COVID-19 symptoms and treatments. Given that the online information ecosystem is extremely noisy, fraught with misinformation and disinformation, and often contaminated by malicious agents spreading propaganda, identifying genuine and good quality information from disinformation is a challenging task for humans. In this regard, there is a significant amount of ongoing research in the directions of disinformation detection and mitigation. In this survey, we discuss the online disinformation problem, focusing on the recent ″infodemic″ in the wake of the coronavirus pandemic. We then proceed to discuss the inherent challenges in disinformation research, including data collection, early detection and effective mitigation, fact-checking based approaches, multi-modality approaches, and policy issues and fairness, and elaborate on the interdisciplinary approaches towards the detection and mitigation of disinformation, after a short overview of the various directions explored in computational detection and mitigation efforts.
2021, 58(7): 1366-1384.
DOI: 10.7544/issn1000-1239.2021.20200818
Abstract：
The outbreak of the COVID-19 pandemic is accompanied by numerous rumors spreading on the social media platform, which seriously affects the stability of society and the safety of public. Existing quantitative analyses of COVID-19 related social media rumors only focus on single element of communication, such as content, while ignoring other basic elements of communication, including communicator, audience, and effect. Besides, compared with the real social media rumor data, the rumor data of these studies have distribution bias and lack of information. Therefore, we conduct a more comprehensive quantitative analysis on the communication of COVID-19 related social media rumors based on the Sina Weibo platform. Specifically, we first analyze the communication content of rumors, including the analysis of the topic, involved regions, event tendency and sentiment. Further, we investigate the users engaged in rumor communication and divide the users into three categories, namely, rumor posters, rumor spreaders, and rumor informers. We explore the basic attributes, topic preferences, individual sentiments, and self-network characteristics of the engaged users. Finally, we study the public opinion triggered by rumors, including the overall sentiment distribution, its correlation with topics, keywords and regions, as well as the evolution of sentiment. To conclude, this study first quantitatively analyzes COVID-19 related social media rumors from the perspective of different basic elements in communication. It provides a more comprehensive and profound understanding of COVID-19 related social media rumors and is of great value for both research and management of rumor in public emergencies.
The outbreak of the COVID-19 pandemic is accompanied by numerous rumors spreading on the social media platform, which seriously affects the stability of society and the safety of public. Existing quantitative analyses of COVID-19 related social media rumors only focus on single element of communication, such as content, while ignoring other basic elements of communication, including communicator, audience, and effect. Besides, compared with the real social media rumor data, the rumor data of these studies have distribution bias and lack of information. Therefore, we conduct a more comprehensive quantitative analysis on the communication of COVID-19 related social media rumors based on the Sina Weibo platform. Specifically, we first analyze the communication content of rumors, including the analysis of the topic, involved regions, event tendency and sentiment. Further, we investigate the users engaged in rumor communication and divide the users into three categories, namely, rumor posters, rumor spreaders, and rumor informers. We explore the basic attributes, topic preferences, individual sentiments, and self-network characteristics of the engaged users. Finally, we study the public opinion triggered by rumors, including the overall sentiment distribution, its correlation with topics, keywords and regions, as well as the evolution of sentiment. To conclude, this study first quantitatively analyzes COVID-19 related social media rumors from the perspective of different basic elements in communication. It provides a more comprehensive and profound understanding of COVID-19 related social media rumors and is of great value for both research and management of rumor in public emergencies.
2021, 58(7): 1385-1394.
DOI: 10.7544/issn1000-1239.2021.20200817
Abstract：
Product review information is an important basis for users’ online decision-making. However, driven by profit, businesses often hire professional writers to write a large number of fake reviews to mislead users and achieve the purpose of packaging themselves and denigrating competitors, resulting in unfair business competition and extremely poor user experience. In response to this phenomenon, we improved the existing spam review recognition methods through Pre-training Models, and proposed a joint pre-training learning method that can simultaneously integrate the semantic and sentimental information of product reviews. In view of the powerful semantic representation capabilities of the pre-trained model, we apply two pre-trained encoders to extract the semantic and emotional features of reviews in the joint learning framework. We integrate the two types of features through joint pre-training learning method. Apart from that, we add the Center Loss function to optimize the model. We have conducted several verification experiments on multiple public data sets and multiple different tasks. The experiments show that our proposed joint model has achieved the best results and has a stronger generalization in both fake review detection and sentiment analysis tasks.
Product review information is an important basis for users’ online decision-making. However, driven by profit, businesses often hire professional writers to write a large number of fake reviews to mislead users and achieve the purpose of packaging themselves and denigrating competitors, resulting in unfair business competition and extremely poor user experience. In response to this phenomenon, we improved the existing spam review recognition methods through Pre-training Models, and proposed a joint pre-training learning method that can simultaneously integrate the semantic and sentimental information of product reviews. In view of the powerful semantic representation capabilities of the pre-trained model, we apply two pre-trained encoders to extract the semantic and emotional features of reviews in the joint learning framework. We integrate the two types of features through joint pre-training learning method. Apart from that, we add the Center Loss function to optimize the model. We have conducted several verification experiments on multiple public data sets and multiple different tasks. The experiments show that our proposed joint model has achieved the best results and has a stronger generalization in both fake review detection and sentiment analysis tasks.
2021, 58(7): 1395-1411.
DOI: 10.7544/issn1000-1239.2021.20200810
Abstract：
Rumor detection has aroused increasing attention due to the damage caused by rampant rumors. Recent studies leverage the post content and propagation structure to detect rumors based on deep-learning based models. However, most of them focus on explicit interactions between posts in the spreading process, while ignoring the modeling of implicit relations, which limits the ability to encode the propagation structure. For example, in the interactive pattern of forwarding (commenting), there are often local implicit interactions between multiple forwarded (commented) posts. In this paper, we propose a rumor detection approach based on a multi-relational propagation tree, investigating multiple kinds of dependencies between posts and enhancing the influence of important posts, to capture richer propagation patterns. Specifically, we formulate the textual content and the propagation tree structure as a heterogeneous graph. Then, we present a novel multi-relational graph convolutional network to learn the inter-level dependency between parent and child nodes and the intra-level dependency between sibling nodes. Meanwhile, we exploit the source post and key spreading posts to model the potential influence of important posts in the spreading process. Finally, we aggregate the node features to learn a more discriminative representation for rumor detection. Extensive experiments on three public benchmark datasets demonstrate the effectiveness and superiority of our proposed method.
Rumor detection has aroused increasing attention due to the damage caused by rampant rumors. Recent studies leverage the post content and propagation structure to detect rumors based on deep-learning based models. However, most of them focus on explicit interactions between posts in the spreading process, while ignoring the modeling of implicit relations, which limits the ability to encode the propagation structure. For example, in the interactive pattern of forwarding (commenting), there are often local implicit interactions between multiple forwarded (commented) posts. In this paper, we propose a rumor detection approach based on a multi-relational propagation tree, investigating multiple kinds of dependencies between posts and enhancing the influence of important posts, to capture richer propagation patterns. Specifically, we formulate the textual content and the propagation tree structure as a heterogeneous graph. Then, we present a novel multi-relational graph convolutional network to learn the inter-level dependency between parent and child nodes and the intra-level dependency between sibling nodes. Meanwhile, we exploit the source post and key spreading posts to model the potential influence of important posts in the spreading process. Finally, we aggregate the node features to learn a more discriminative representation for rumor detection. Extensive experiments on three public benchmark datasets demonstrate the effectiveness and superiority of our proposed method.
2021, 58(7): 1412-1424.
DOI: 10.7544/issn1000-1239.2021.20200801
Abstract：
Social media not only brings convenience to people, but also provides a platform for spreading rumors. Currently, most rumor detection methods are based on text content information. However, in social media scenarios, text content is mostly short text, which often leads to poor performance due to data sparsity. Message propagation on social networks can be modeled as a graph structure. Previous studies have taken into account the characteristics of message propagation structure, and detected rumors through GCN. GCN aggregates neighbors based on structural information to enhance node representation, but some neighbor aggregation is useless and may even cause noise, which making the representation obtained from GCN unreliable. Meanwhile, these methods can not effectively highlight the importance of the source post information. In this paper, we propose a propagation graph convolution network model GUCNH. In GUCNH model, information forwarding graph is constructed first, and the representation of neighbor nodes is aggregated by two fusion gated convolution network modules. Fusion gating can select and combine the feature representation before and after the graph convolution to get a more reliable representation. Considering that in forwarding graph, any post may interact with each other rather than just with its neighbors, a multi-headed self-attention module is introduced between two integrated gated convolution network modules to model the multi-angle influence between posts. In addition, in forwarding graph, the source posts often contain the richest information than reposts. After generating each node representation, we selectively enhance the source node’s information to enhance the influence of the source posts. Experiments on three real datasets show that our proposed model outperforms the existing methods.
Social media not only brings convenience to people, but also provides a platform for spreading rumors. Currently, most rumor detection methods are based on text content information. However, in social media scenarios, text content is mostly short text, which often leads to poor performance due to data sparsity. Message propagation on social networks can be modeled as a graph structure. Previous studies have taken into account the characteristics of message propagation structure, and detected rumors through GCN. GCN aggregates neighbors based on structural information to enhance node representation, but some neighbor aggregation is useless and may even cause noise, which making the representation obtained from GCN unreliable. Meanwhile, these methods can not effectively highlight the importance of the source post information. In this paper, we propose a propagation graph convolution network model GUCNH. In GUCNH model, information forwarding graph is constructed first, and the representation of neighbor nodes is aggregated by two fusion gated convolution network modules. Fusion gating can select and combine the feature representation before and after the graph convolution to get a more reliable representation. Considering that in forwarding graph, any post may interact with each other rather than just with its neighbors, a multi-headed self-attention module is introduced between two integrated gated convolution network modules to model the multi-angle influence between posts. In addition, in forwarding graph, the source posts often contain the richest information than reposts. After generating each node representation, we selectively enhance the source node’s information to enhance the influence of the source posts. Experiments on three real datasets show that our proposed model outperforms the existing methods.
2021, 58(7): 1425-1435.
DOI: 10.7544/issn1000-1239.2021.20200806
Abstract：
In online social networks, massive amounts of information are transmitted and diffused through users’ interaction and reposting behavior. As the carrier of information diffusion, social media can not only make people share information flow and get current affairs news quickly, but also facilitate the exchange of ideas and information between people. At the same time, it may become an important channel for the spread of false information. Most of the existing researches on false information detection are based on the recognition models of machine learning and deep learning of Weibo content, while ignoring the structural differences between true and false information networks. Therefore, based on the motif theory of complex networks, this paper puts forward the concepts of breadth and depth motif degree to quantify the structural importance of the network. The research shows that the importance calculation method based on motif degree is an innovation and expansion of traditional network structure importance index, which can measure the specificity of communication network structure more comprehensively. This paper analyzes and reveals the structure characteristics and propagation mechanism of false information in microblog network by constructing the two-dimensional motif measurement index, that is, the false information is diffused under the joint action of breadth and depth propagation, and the breadth motif mainly affects the network spread scale, while the depth motif degree affects the complexity of the network structure. Even in the early stage of information diffusion, the false news detection method based on motif features has a high prediction accuracy. The network feature analysis based on motif degree can be applied to detect false information from the source in the early stage of social media information diffusion, which provides a novel and feasible way for false information detection.
In online social networks, massive amounts of information are transmitted and diffused through users’ interaction and reposting behavior. As the carrier of information diffusion, social media can not only make people share information flow and get current affairs news quickly, but also facilitate the exchange of ideas and information between people. At the same time, it may become an important channel for the spread of false information. Most of the existing researches on false information detection are based on the recognition models of machine learning and deep learning of Weibo content, while ignoring the structural differences between true and false information networks. Therefore, based on the motif theory of complex networks, this paper puts forward the concepts of breadth and depth motif degree to quantify the structural importance of the network. The research shows that the importance calculation method based on motif degree is an innovation and expansion of traditional network structure importance index, which can measure the specificity of communication network structure more comprehensively. This paper analyzes and reveals the structure characteristics and propagation mechanism of false information in microblog network by constructing the two-dimensional motif measurement index, that is, the false information is diffused under the joint action of breadth and depth propagation, and the breadth motif mainly affects the network spread scale, while the depth motif degree affects the complexity of the network structure. Even in the early stage of information diffusion, the false news detection method based on motif features has a high prediction accuracy. The network feature analysis based on motif degree can be applied to detect false information from the source in the early stage of social media information diffusion, which provides a novel and feasible way for false information detection.
2021, 58(7): 1436-1455.
DOI: 10.7544/issn1000-1239.2021.20200685
Abstract：
Graph anomaly detection aims to find “strange” or “unusual” patterns in large graph or massive graph databases, and it has a wide range of application scenarios. Deep learning can learn the hidden rules from the data, and it has excellent performance in extracting potential complex patterns from data. With the great development of graph representation learning in recent years, how to detect graph anomaly using deep learning methods has attracted extensive attention in the area of academia and industry. Although a series of recent studies have investigated anomaly detection methods from the perspective of graphs, there is a lack of attention to graph anomaly detection methods under the background of deep learning. In this paper, we first give the definitions of various kinds of anomalies in static graph and dynamic graph and investigate the deep neural network based graph representation learning method and its various applications in graph anomaly detection. Then we present the current situation of research on graph anomaly detection based on deep learning from the perspective of static graph and dynamic graph, and summarize the application scenarios and related data sets of graph anomaly detection. At last, we discuss the current challenges and future research directions of graph anomaly detection.
Graph anomaly detection aims to find “strange” or “unusual” patterns in large graph or massive graph databases, and it has a wide range of application scenarios. Deep learning can learn the hidden rules from the data, and it has excellent performance in extracting potential complex patterns from data. With the great development of graph representation learning in recent years, how to detect graph anomaly using deep learning methods has attracted extensive attention in the area of academia and industry. Although a series of recent studies have investigated anomaly detection methods from the perspective of graphs, there is a lack of attention to graph anomaly detection methods under the background of deep learning. In this paper, we first give the definitions of various kinds of anomalies in static graph and dynamic graph and investigate the deep neural network based graph representation learning method and its various applications in graph anomaly detection. Then we present the current situation of research on graph anomaly detection based on deep learning from the perspective of static graph and dynamic graph, and summarize the application scenarios and related data sets of graph anomaly detection. At last, we discuss the current challenges and future research directions of graph anomaly detection.
2021, 58(7): 1456-1465.
DOI: 10.7544/issn1000-1239.2021.20200804
Abstract：
In recent years, social media has become the main access where people acquire the latest news. However, the convenience and openness of social media have also facilitated the proliferation of fake news. With the development of multimedia technology, fake news on social media has been evolving from text-only posts to multimedia posts containing images or videos. Therefore, multi-modal fake news detection is attracting more and more attention. Existing methods for multi-modal fake news detection mostly focus on capturing appearance-level features that are highly dependent on the dataset distribution but insufficiently exploit the semantics-level features. Thus, the methods often fail to understand the deep semantics of textual and visual entities in the fake news, which indeed limits the generalizability of models in real applications. To tackle this problem, this paper proposes a semantics-enhanced multi-modal model for fake news detection, which better models the underlying semantics of multi-modal news by implicitly utilizing the factual knowledge in the pre-trained language model and explicitly extracting the visual entities. Furthermore, the proposed method extracts visual features of different semantic levels and models the semantic interaction between the textual and visual features by the text-guided attention mechanism, which better fuses the multi-modal heterogeneous features. Extensive experiments on the Weibo dataset strongly evidence that our method outperforms the state of the art significantly.
In recent years, social media has become the main access where people acquire the latest news. However, the convenience and openness of social media have also facilitated the proliferation of fake news. With the development of multimedia technology, fake news on social media has been evolving from text-only posts to multimedia posts containing images or videos. Therefore, multi-modal fake news detection is attracting more and more attention. Existing methods for multi-modal fake news detection mostly focus on capturing appearance-level features that are highly dependent on the dataset distribution but insufficiently exploit the semantics-level features. Thus, the methods often fail to understand the deep semantics of textual and visual entities in the fake news, which indeed limits the generalizability of models in real applications. To tackle this problem, this paper proposes a semantics-enhanced multi-modal model for fake news detection, which better models the underlying semantics of multi-modal news by implicitly utilizing the factual knowledge in the pre-trained language model and explicitly extracting the visual entities. Furthermore, the proposed method extracts visual features of different semantic levels and models the semantic interaction between the textual and visual features by the text-guided attention mechanism, which better fuses the multi-modal heterogeneous features. Extensive experiments on the Weibo dataset strongly evidence that our method outperforms the state of the art significantly.
2021, 58(7): 1466-1475.
DOI: 10.7544/issn1000-1239.2021.20200799
Abstract：
Audio deepfake detection is a hot topic in recent years and has been widely concerned. At present, convolutional neural networks and their variants have made good progress in the task of audio deepfake detection. However, there are still two problems: 1) The assumption of current work is that each aspect of the feature map fed into the convolutional neural network has the same effect on the result, ignoring that the information emphasized at different locations on each dimensional feature map is different. 2) In addition, the current work focuses on the local information of the feature map, and cannot make use of the relationship between the feature map from the global view. To solve these challenges, we introduce a global and temporal-frequency attention based network that focuses on channel dimensions and temporal-frequency dimensions, respectively. Specifically, we introduced two parallel attention modules. One is the temporal-frequency attention module and the other is the global attention module. For the temporal-frequency attention module, we can update the features by using weighted aggregation on all temporal-frequency feature maps. For the global attention module, we draw on the idea of SE-Net to generate weights for each feature channel by parameters. And by this way, we can get the global distribution of the response on the feature channel. We have carried out a series of experiments on ASVspoof2019 LA open data set, and the results showed that the proposed model achieved good results, and the EER of the best model reached 4.12%, which refreshed the best results of the single model.
Audio deepfake detection is a hot topic in recent years and has been widely concerned. At present, convolutional neural networks and their variants have made good progress in the task of audio deepfake detection. However, there are still two problems: 1) The assumption of current work is that each aspect of the feature map fed into the convolutional neural network has the same effect on the result, ignoring that the information emphasized at different locations on each dimensional feature map is different. 2) In addition, the current work focuses on the local information of the feature map, and cannot make use of the relationship between the feature map from the global view. To solve these challenges, we introduce a global and temporal-frequency attention based network that focuses on channel dimensions and temporal-frequency dimensions, respectively. Specifically, we introduced two parallel attention modules. One is the temporal-frequency attention module and the other is the global attention module. For the temporal-frequency attention module, we can update the features by using weighted aggregation on all temporal-frequency feature maps. For the global attention module, we draw on the idea of SE-Net to generate weights for each feature channel by parameters. And by this way, we can get the global distribution of the response on the feature channel. We have carried out a series of experiments on ASVspoof2019 LA open data set, and the results showed that the proposed model achieved good results, and the EER of the best model reached 4.12%, which refreshed the best results of the single model.
2021, 58(7): 1476-1489.
DOI: 10.7544/issn1000-1239.2021.20200803
Abstract：
With the rapid development of generative adversarial networks (GAN), breakthrough progress has been made in fake face generation. In order to reduce the harmful effects of fake face generation technology to society, fake face identification has become a very important topic, which has attracted numerous research efforts. Although impressive progress has been made in fake face identification, there are still many problems to be solved. Among them, how to improve the generalization ability of the fake face detection model is a crucial issue, and it is also the key to deploy fake face detection techniques in real-world scenarios. This paper studies the fake face identification problem, aiming to improve the generalization ability of the fake face identification model. Motivated by the idea of unsupervised domain adaptation, this paper introduces the domain adversarial branch to weaken the extraction of non-robust features of specific generative models by the feature extractor, so that the model can extract features with stronger robustness and higher generalization ability, improving the identification performance on the fake face images generated by unknown GANs. Experimental results show that the method proposed in this paper can effectively improve the generalization ability of the identification model, and significantly improve the performance of the fake face identification model on the fake images generated by the unknown generation model.
With the rapid development of generative adversarial networks (GAN), breakthrough progress has been made in fake face generation. In order to reduce the harmful effects of fake face generation technology to society, fake face identification has become a very important topic, which has attracted numerous research efforts. Although impressive progress has been made in fake face identification, there are still many problems to be solved. Among them, how to improve the generalization ability of the fake face detection model is a crucial issue, and it is also the key to deploy fake face detection techniques in real-world scenarios. This paper studies the fake face identification problem, aiming to improve the generalization ability of the fake face identification model. Motivated by the idea of unsupervised domain adaptation, this paper introduces the domain adversarial branch to weaken the extraction of non-robust features of specific generative models by the feature extractor, so that the model can extract features with stronger robustness and higher generalization ability, improving the identification performance on the fake face images generated by unknown GANs. Experimental results show that the method proposed in this paper can effectively improve the generalization ability of the identification model, and significantly improve the performance of the fake face identification model on the fake images generated by the unknown generation model.
2021, 58(7): 1490-1503.
DOI: 10.7544/issn1000-1239.2021.20200809
Abstract：
In recent years, with the increasing number of users in social networks such as Twitter, Facebook and Sina Weibo, the amount of information has rapidly expanded. The spread of untrue information hidden in massive information has brought adverse effects. How to regulate or suppress the spread of specific information is a technical challenge faced by network information management. In order to solve this problem, first of all, the independent information forwarding prediction mechanism based on machine learning method, which does not depend on the propagation model is proposed, so as to predict the information propagation. Secondly, based on the independent cascade model, considering the particularity of the scenario in this paper, the asynchronous information unequal competition propagation model is proposed as the competitive propagation mechanism of specific information and immune information. Finally, three selection algorithms of seed nodes are proposed and the immune information is widely spread in the network by injecting immune information into the seed nodes, so as to suppress the spread of specific information. Experiments based on real social network data show that the information propagation prediction model and the selection algorithms of seed nodes proposed have good effects on the regulation and suppression of specific information propagation.
In recent years, with the increasing number of users in social networks such as Twitter, Facebook and Sina Weibo, the amount of information has rapidly expanded. The spread of untrue information hidden in massive information has brought adverse effects. How to regulate or suppress the spread of specific information is a technical challenge faced by network information management. In order to solve this problem, first of all, the independent information forwarding prediction mechanism based on machine learning method, which does not depend on the propagation model is proposed, so as to predict the information propagation. Secondly, based on the independent cascade model, considering the particularity of the scenario in this paper, the asynchronous information unequal competition propagation model is proposed as the competitive propagation mechanism of specific information and immune information. Finally, three selection algorithms of seed nodes are proposed and the immune information is widely spread in the network by injecting immune information into the seed nodes, so as to suppress the spread of specific information. Experiments based on real social network data show that the information propagation prediction model and the selection algorithms of seed nodes proposed have good effects on the regulation and suppression of specific information propagation.
2021, 58(7): 1504-1517.
DOI: 10.7544/issn1000-1239.2021.20200112
Abstract：
Convolutional neural network (CNN) achieves good performance in image processing, speech recognition, natural language processing and other fields. Large-scale neural network models often encounter resource constraints such as computing and storage. The emergence of sparse neural networks effectively relieves the need for computing and storage. Although existing domain-specific accelerators can effectively handle sparse networks, they achieve high energy efficiency through tight coupling of algorithms and structures, and lose the flexibility of the structure. The coarse-grained dataflow architecture can implement different neural network applications through flexible instruction scheduling. Based on this architecture, the regular computing characteristics of dense convolution allow different channels to share the same set of instruction to execute. However, there are sparse weights in sparse networks, making these instructions have 0-value-related invalid instructions, which makes the existing instruction execution method cannot automatically skip them, resulting in invalid calculations. At the same time, when executing an irregular sparse network, existing instruction mapping methods cause an unbalanced load on the computing array. These problems hinder the improvement of sparse network performance. In this paper, based on the premise that different channels share a set of instructions, we add an instruction control unit based on the data and instruction characteristics of the sparse network to achieve detection and skipping of 0-value related instructions in the weight data, while using the load balanced instruction mapping algorithm to solve the problem of uneven instruction execution in sparse networks. Experiments show that compared with dense networks, sparse networks achieve an average performance increase of 1.55X and an energy reduction of 63.77%. In addition, it achieves 2.39X(Alexnet), 2.28X(VGG16) and 1.14X(Alexnet), 1.23X(VGG16) speedup over GPU (cuSparse) and Cambricon-X, respectively.
Convolutional neural network (CNN) achieves good performance in image processing, speech recognition, natural language processing and other fields. Large-scale neural network models often encounter resource constraints such as computing and storage. The emergence of sparse neural networks effectively relieves the need for computing and storage. Although existing domain-specific accelerators can effectively handle sparse networks, they achieve high energy efficiency through tight coupling of algorithms and structures, and lose the flexibility of the structure. The coarse-grained dataflow architecture can implement different neural network applications through flexible instruction scheduling. Based on this architecture, the regular computing characteristics of dense convolution allow different channels to share the same set of instruction to execute. However, there are sparse weights in sparse networks, making these instructions have 0-value-related invalid instructions, which makes the existing instruction execution method cannot automatically skip them, resulting in invalid calculations. At the same time, when executing an irregular sparse network, existing instruction mapping methods cause an unbalanced load on the computing array. These problems hinder the improvement of sparse network performance. In this paper, based on the premise that different channels share a set of instructions, we add an instruction control unit based on the data and instruction characteristics of the sparse network to achieve detection and skipping of 0-value related instructions in the weight data, while using the load balanced instruction mapping algorithm to solve the problem of uneven instruction execution in sparse networks. Experiments show that compared with dense networks, sparse networks achieve an average performance increase of 1.55X and an energy reduction of 63.77%. In addition, it achieves 2.39X(Alexnet), 2.28X(VGG16) and 1.14X(Alexnet), 1.23X(VGG16) speedup over GPU (cuSparse) and Cambricon-X, respectively.
2021, 58(7): 1518-1532.
DOI: 10.7544/issn1000-1239.2021.20200690
Abstract：
Flash memory based solid state drives (SSDs) are widely deployed in personal computers, data centers, and cloud storages given their well identified advantages, such as high performance, low power consumption and non-volatile property. In recent years, with the development of process technology and micro electronic technology, the features of SSDs are greatly changed. First, detailed storage mechanism of flash cell is introduced, including architecture of flash cell and flash block, programming method, and basic operation of SSDs. Then, several SSD controller key technologies are presented, including buffer device, flash translation layer, garbage collection, data allocation, wear leveling and error correction code. These technologies are used to well support normal operation of SSDs. Additionally, the parallel architecture of SSDs which is used to boost the performance of SSDs, is discussed and related constraints are also presented while several previous works on parallelism exploration are analyzed. Next, since the scaling of SSD has evolved from planar (2D) to 3D stacking, 3D SSD is introduced as a new type of SSDs that can provide larger capacity compared with traditional planar SSD. In this paper, the characteristics of 3D SSDs’ vertical architecture, performance and lifetime are analyzed. Also, the disadvantages of previous works on 3D SSD performance and lifetime optimizations are discussed. Finally, current state of SSDs is summarized and possible future research works are given.
Flash memory based solid state drives (SSDs) are widely deployed in personal computers, data centers, and cloud storages given their well identified advantages, such as high performance, low power consumption and non-volatile property. In recent years, with the development of process technology and micro electronic technology, the features of SSDs are greatly changed. First, detailed storage mechanism of flash cell is introduced, including architecture of flash cell and flash block, programming method, and basic operation of SSDs. Then, several SSD controller key technologies are presented, including buffer device, flash translation layer, garbage collection, data allocation, wear leveling and error correction code. These technologies are used to well support normal operation of SSDs. Additionally, the parallel architecture of SSDs which is used to boost the performance of SSDs, is discussed and related constraints are also presented while several previous works on parallelism exploration are analyzed. Next, since the scaling of SSD has evolved from planar (2D) to 3D stacking, 3D SSD is introduced as a new type of SSDs that can provide larger capacity compared with traditional planar SSD. In this paper, the characteristics of 3D SSDs’ vertical architecture, performance and lifetime are analyzed. Also, the disadvantages of previous works on 3D SSD performance and lifetime optimizations are discussed. Finally, current state of SSDs is summarized and possible future research works are given.
2021, 58(7): 1533-1543.
DOI: 10.7544/issn1000-1239.2021.20200672
Abstract：
Industrial applications usually have strict requirements of data transmission certainty. It is therefore essential for industrial edge computing applications to deploy a proper caching strategy at edge nodes, in order to ensure the real-time performance guarantee. The cache optimization problem is formulized considering the specific requirements of industrial applications. The content request is modeled as shot noise model (SNM) to reflect the dynamic characteristics of popularity. A scheme of popularity prediction is then proposed by defining a feature similarity function of the requested content set in the latest periodic time window. Based on it, a new cache replacement algorithm called combing periodic popularity prediction and size caching strategy (PPPS) is proposed. The value of each cache content is determined together with the popularity, size and time updates parameters. The content with minimum value will be deleted with the highest priority when content replacement happens. The experimental results show that the proposed PPPS algorithm outperforms all the 5 baseline algorithms, which are the most popular content (MPC), greedy dual size (GDS), least recently used (LRU), least frequently used (LFU), and FIFO algorithm. PPPS algorithm obtains the best performance of hit rate and the average delay in all the testing cases with different parameter settings on user request models, content size distributions, and content types.
Industrial applications usually have strict requirements of data transmission certainty. It is therefore essential for industrial edge computing applications to deploy a proper caching strategy at edge nodes, in order to ensure the real-time performance guarantee. The cache optimization problem is formulized considering the specific requirements of industrial applications. The content request is modeled as shot noise model (SNM) to reflect the dynamic characteristics of popularity. A scheme of popularity prediction is then proposed by defining a feature similarity function of the requested content set in the latest periodic time window. Based on it, a new cache replacement algorithm called combing periodic popularity prediction and size caching strategy (PPPS) is proposed. The value of each cache content is determined together with the popularity, size and time updates parameters. The content with minimum value will be deleted with the highest priority when content replacement happens. The experimental results show that the proposed PPPS algorithm outperforms all the 5 baseline algorithms, which are the most popular content (MPC), greedy dual size (GDS), least recently used (LRU), least frequently used (LFU), and FIFO algorithm. PPPS algorithm obtains the best performance of hit rate and the average delay in all the testing cases with different parameter settings on user request models, content size distributions, and content types.
2021, 58(7): 1544-1562.
DOI: 10.7544/issn1000-1239.2021.20200480
Abstract：
The features of software defined networking (SDN) such as forwarding and control separation, centralized control, and open interfaces make the network flexible and controllable, and its architecture has been fully developed. Due to the good combination with various cloud services, SDN has received a large number of commercial deployments in recent years. In OpenFlow-based SDN architecture, ternary content addressable memory (TCAM) is mostly used on hardware switches to store flow entries installed by the controller in order to achieve such goals as fast lookup of flow entries and mask matching. However, limited by the capacity and price of TCAM, the current commercial OpenFlow switches can store at most tens of thousands of flow entries, which leads to the problem of flow table overflow caused by burst traffic or flow table overflow attacks, which seriously affects the network performance. How to establish an efficient flow table overflow mitigation mechanism has attracted extensive attention from researchers. Firstly, the causes and effects of flow table overflow problem in OpenFlow switch are discussed. On this basis, the current research status of flow table overflow mitigation technology is summarized and compared according to the two situations of burst traffic and attack behavior. Finally, the existing research problems are summarized and analyzed, and the future development direction and challenges are forecasted.
The features of software defined networking (SDN) such as forwarding and control separation, centralized control, and open interfaces make the network flexible and controllable, and its architecture has been fully developed. Due to the good combination with various cloud services, SDN has received a large number of commercial deployments in recent years. In OpenFlow-based SDN architecture, ternary content addressable memory (TCAM) is mostly used on hardware switches to store flow entries installed by the controller in order to achieve such goals as fast lookup of flow entries and mask matching. However, limited by the capacity and price of TCAM, the current commercial OpenFlow switches can store at most tens of thousands of flow entries, which leads to the problem of flow table overflow caused by burst traffic or flow table overflow attacks, which seriously affects the network performance. How to establish an efficient flow table overflow mitigation mechanism has attracted extensive attention from researchers. Firstly, the causes and effects of flow table overflow problem in OpenFlow switch are discussed. On this basis, the current research status of flow table overflow mitigation technology is summarized and compared according to the two situations of burst traffic and attack behavior. Finally, the existing research problems are summarized and analyzed, and the future development direction and challenges are forecasted.
2021, 58(7): 1563-1572.
DOI: 10.7544/issn1000-1239.2021.20200018
Abstract：
Computer networks have been playing an important role in modern society. The rapid growth of the network scale makes the network traffic more and more complicated, which is hard to accurately model. This condition makes the optimal routing policy in communication networks an NP-hard problem. To solve this problem, traditional methods for routing and traffic engineering mainly use hand-crafted algorithms, which cannot ensure both the accuracy and efficiency. In recent years, deep reinforcement learning (DRL)-based network routing strategies have been proposed, which overcome the shortcomings of manually analysis and modelling by human experts to some extent. However, current DRL-based routing strategies all have problems in scalability, which means they cannot be used in large scale networks. Under this circumstance, this paper proposes Hierar-DRL, a DRL-based network routing technology that employs pinning control theory. Pinning control helps Hierar-DRL to select a subset of network nodes as the target control nodes of DRL. With the advantages of pinning control and the automatic policy exploring ability of DRL, Hierar-DRL shows better scalability in large networks. Simulation results show that the proposed scheme can reduce the average end-to-end transmission delay in the test network topologies by up to 28.5% compared with the state-of-the-art, which validates the proposed scheme.
Computer networks have been playing an important role in modern society. The rapid growth of the network scale makes the network traffic more and more complicated, which is hard to accurately model. This condition makes the optimal routing policy in communication networks an NP-hard problem. To solve this problem, traditional methods for routing and traffic engineering mainly use hand-crafted algorithms, which cannot ensure both the accuracy and efficiency. In recent years, deep reinforcement learning (DRL)-based network routing strategies have been proposed, which overcome the shortcomings of manually analysis and modelling by human experts to some extent. However, current DRL-based routing strategies all have problems in scalability, which means they cannot be used in large scale networks. Under this circumstance, this paper proposes Hierar-DRL, a DRL-based network routing technology that employs pinning control theory. Pinning control helps Hierar-DRL to select a subset of network nodes as the target control nodes of DRL. With the advantages of pinning control and the automatic policy exploring ability of DRL, Hierar-DRL shows better scalability in large networks. Simulation results show that the proposed scheme can reduce the average end-to-end transmission delay in the test network topologies by up to 28.5% compared with the state-of-the-art, which validates the proposed scheme.
