Vol. 41 · No. 6

There are still two problems of the existing methods of defending against poisoning attacks of the blockchain-based federated learning: 1) It is difficult to accurately identify the nodes under attack; 2) The effect of the model is greatly affected when the number of malicious nodes exceeds a half. So, an innovative secure mechanism is proposed for blockchain-based federated learning, which is called the training behavior verification mechanism. The mechanism describes the consistent training behavior rules of nodes by constructing the training behavior model, and distinguishes honest nodes from malicious nodes by comparing the differences in training behavior models on the training behavior verification algorithm. Experiments show that the new mechanism can effectively resist more than half of the label-flipping attacks and backdoor attacks, and has the advantages of higher stability and higher accuracy than methods such as Krum, Trimmed Mean, and Median.

Internet of Things (IoT) has been recognised as a promising area for automating numerous processes, however, the major problem with IoT is its potential for rising complexities. Several approaches have moved attention to the edge nodes associated with IoT, hence concepts of edge-computing, resource allocation and load balancing are tantamount to a more robust heterogeneous IoT. The resource optimization terrain comes with several complications for the resource allocation and scheduling algorithms. Load balancing, one of the key strategies for improving system performance and resource utilization in distributed and parallel computing, generally views an effective load balancer as a `traffic controller' of resources by directing tasks to available and capable resources. In this paper, a framework appropriate for modelling and reasoning about IoT resource optimization is developed. Further, implementation of an optimized resource allocation algorithm taking into consideration the users' quality of experience (QoE) and the quality of service (QoS) is made available. Simulation results authenticate analysis and validate the improved performance over existing work.

In order to eliminate the hidden dangers caused by tire bubble defects, considering that the two-dimensional technology is sensitive to light, the 3D point cloud technology is used to obtain the tire surface morphology. This paper proposes a 3D point cloud network model named PointVotes, a point based target detection method. The designed structural framework includes: the fusion sampling layer, the voting layer and the proposal refinement layer. By observing the spatial characteristics of the detected target, a new point sampling method named C-farthest point sampling (C-FPS) is proposed. Combining with the fusion sampling strategy, the FPS and the C-FPS are sampled in a certain proportion. It solves the problem that the proposal box cannot be generated due to less available prospect information when generating suggestions for small targets. The network model uses Set Abstraction layers in multiple PointNet++ to extract features, arranges and combines features of different scales, forms high-dimensional features of points and votes, judges whether there are bubble defects through classification, and then generates proposals and regression to the prediction frame. Experiment results show that the mean average precision of the model can reach 82.8

The scale of data shows an explosive growth trend, with wide use of cloud storage. However, there are challenges such as network latency and energy consumption. The emergence of edge computing brings data close to the edge of the network, making it a good supplement to cloud computing. The spatiotemporal characteristics of data have been largely ignored in studies of data placement and storage optimization. To this end, a temperature matrix-based data placement method using an improved Hungarian algorithm (TEMPLIH) is proposed in this work. A temperature matrix is used to reflect the influence of data characteristics on its placement. A data replica matrix selection algorithm based on temperature matrix (RSA-TM) is proposed to meet latency requirements. Then, an improved Hungarian algorithm based on replica matrix (IHA-RM) is proposed, which satisfies the balance among the multiple goals of latency, cost, and load balancing. Compared with other data placement strategies, experiments show that the proposed method can effectively reduce the cost of data placement while meeting user access latency requirements and maintaining a reasonable load balance between edge servers. Further improvement is discussed and the idea of regional value is proposed.

As social networks have been rapidly growing, traditional network representation learning methods are struggling to accurately characterize their dynamic changes, and to output effective node classification and link prediction. To address this problem, this paper proposes IproGRU, a dynamic network representation learning method based on an improved Gated Recurrent Unit (GRU) network to improve the dynamic network representation. First, the method quickly generates embedding for an influenced node by sampling and aggregating features of its neighboring nodes when the network changes. Second, it updates the embedding of the influenced node on time series by the improved GRU network to fully adapt to the changes of the dynamic network. Experimental results on node classification and link prediction for three datasets of dynamic networks show that the proposed method improves the accuracy by 5–10

For the study of information dissemination in online social networks, most existing information dissemination models include only positive relationships, ignoring the existence and importance of negative relationships, and do not consider the influence of inter-individual relationship polarity on dissemination. To solve these problems, we propose a social network information dissemination model incorporating negative relationships in this paper. Drawing on the state concept of the SIR (Susceptible Infected Recovered) model, the three types of SIR states are subdivided into five sub-states. Combining the advantages of the viewpoint evolution model, the influence of relational polarity on node attitudes is added to the modeling of the propagation process. The experiment proves that the method proposed in this paper can show more specifically the changing trend in the number of propagation nodes with different attitudes and portray the process of information propagation in online social networks.

Generative Adversarial Networks have gained popularity mainly due to their ability to create fake human faces. The remarkable detail with which such images have been created in the past few years has exceeded the ability of humans to differentiate between these fake images and real images. Such images have been known to be capable of deceiving the face recognition systems with certain success as well. Forensic systems being developed nowadays take into account adversarial attacks in order to create a more comprehensive detection approaches. Different GAN algorithms such as StackGAN, StyleGAN use different architectures to produce images. Since the underlying technique is different from one another it is difficult for any single detection algorithm trained on one kind of GAN to detect fake images generated from some other kind of GAN. In this research we use a siamese network with triplet loss function to provide a generic solution for detection of GAN generated images or deepfake images. Extensive experiments have been conducted to analyze the effectiveness of the proposed approach. The results show that the siamese triplet loss network performs significantly better than the contemporary approaches with accuracy exceeding 90

Network representation learning has attracted widespread attention as a pre-processing process for some machine learning and deep learning tasks. However, most existing methods only consider influence of nodes' low-order neighbors to represent them. Either nodes' high-order neighbor or the intrinsic characteristic attributes of nodes are ignored, leading to the effect of network representation learning that needs to be improved. This paper proposes a novel model named Structure Enhanced Graph Convolutional Network (SEGCN) to address these limitations. SEGCN consists of the following components, i.e., the network structure enhancement to transform weak relationship into strong relationship, the node feature aggregation to fuse high-order neighbor information. Hence, the SEGCN model can simultaneously integrate network structure information, attribute information, and high-order neighbor relationships together. Experimental results for node classification and node clustering on six datasets show that SEGCN achieves better effectiveness and efficiency than state-of-the-art baselines.

Alzheimer's dementia (AD) is characterized by memory loss, which is one of the earliest symptoms to develop. In this study, we investigated audio transcript data of patients with Alzheimer's dementia. The study involved the use of three intelligent computational approaches: conventional machine learning (Support Vector Machine, Random Forest, Decision Tree), sequential deep learning (LSTM, bidirectional LSTM, CNN-LSTM), and transfer learning (BERT, XLNet) models for automatic detection of linguistic indicators for early diagnosis of Alzheimer's dementia. These models were trained on the DementiaBank clinical transcript dataset. The grid search tuning approach is used for tuning the values of the hyperparameters. Text vectorization is done using the Term Frequency-Inverse Document Frequency (TF-IDF) information retrieval approach. TF-IDF is based on the Bag of Words (BoW) paradigm, which deals with the less and more relevant words in a transcript. Results were evaluated and compared using several performance metrics. The state-of-the-art techniques implemented on DementiaBank dataset in our methodology achieved better performance in terms of accuracy. Transfer learning models showed better classification results in comparison to sequential deep learning models. However, sequential deep learning models outperformed traditional machine learning models. Overall, in terms of accuracy, BERT and XLNet were the most accurate, with accuracy of 93

Event Extraction (EE) is an important task in Natural Language Understanding (NLU). As the complexity of Chinese structure, Chinese EE is more difficult than English EE. According to the characteristics of Chinese, this paper designed a Semantic-GRU (Sem-GRU) model, which integrates Chinese word context semantics, Chinese word glyph semantics and Chinese word structure semantics. And this paper uses the model for Chinese Event Trigger Extraction (ETE) task. The experiment is compared in two tasks: ETE and Named Entity Recognition (NER). In ETE, the paper uses ACE 2005 Chinese event dataset to compare the existing research, the effect reaches 75.8