Journal of Ambient Intelligence and Humanized Computing最新文献

Predicting human mobility is essential for urban planning and personalized services. The problem addressed in this study is analyzing user behavior patterns and predicting their next destination. Due to the complexity and diversity of human mobility, it’s necessary to study user behavior patterns from various angles and leverage diverse context information to construct prediction models. Unfortunately, most previous research often neglects personalized preferences and falls short in offering a comprehensive understanding of user behavior patterns. Furthermore, some studies have not effectively mined and utilized contextual information. To address these shortcomings, this paper introduces a novel Personalized Behavior Modeling Network (PBMN). Compared to existing methods, PBMN provides a more comprehensive modeling of user behavior and utilizes context information more extensively, enabling more accurate prediction. It models user behavior through two parallel channels, taking into account both sequential patterns and personalized preferences, while fully utilizing different contextual information. Ultimately, it generates prediction results by personalized integration of different behavior features. Specifically, PBMN employs a pair of attention-based encoders and decoders to model the overall behavior features. Additionally, it utilizes three parallel recurrent neural networks to model recent behavior features at different levels of context information. The performance of PBMN was evaluated using two real-world datasets. Experimental results demonstrate that PBMN outperforms five mainstream prediction methods concerning three commonly used evaluation metrics, emphasizing the effectiveness of PBMN

Recently, deep-learning-based models have achieved impressive performance on tasks that were previously considered to be extremely challenging. However, recent works have shown that various deep learning models are susceptible to adversarial data samples. In this paper, we propose the sliced Wasserstein adversarial training method to encourage the logit distributions of clean and adversarial data to be similar to each other. We capture the dissimilarity between two distributions using the Wasserstein metric and then align distributions using an end-to-end training process. We present the theoretical background of the motivation for our study by providing generalization error bounds for adversarial data samples. We performed experiments on three standard datasets and the results demonstrate that our method is more robust against white box attacks compared to previous methods.

The field of computer networking is experiencing rapid growth, accompanied by the swift advancement of internet tools. As a result, people are becoming more aware of the importance of network security. One of the primary concerns in ensuring security is the authority over domains, and network owners are striving to establish a common language to exchange security information and respond quickly to emerging threats. Given the increasing prevalence of various types of attacks, network security has become a significant challenge in the realm of computing. To address this, a multi-level distributed approach incorporating vulnerability identification, dimensioning, and countermeasures based on attack graphs has been developed. Implementing reconfigurable virtual systems as countermeasures significantly improves attack detection and mitigates the impact of attacks. Password-based authentication, for instance, can be susceptible to password cracking techniques, social engineering attacks, or data breaches that expose user credentials. Similarly, ensuring privacy during data transmission through encryption helps protect data from unauthorized access, but it does not guarantee the prevention of other types of attacks such as malware infiltration or insider threats. This research explores various techniques to achieve effective attack detection. Multiple research methods have been utilized and evaluated to identify the most suitable approach for network security and attack detection in the context of cloud computing. The analysis and implementation of diverse research studies demonstrate that the based signature intrusion detection method outperforms others in terms of precision, recall, F-measure, accuracy, reliability, and time complexity.

In wireless communications networks, the non-orthogonal multiple access (NOMA) technique is different from the existing orthogonal multiple access (OMA) techniques. In NOMA, the available number of resources are more and it leads to multiple access interference. In this paper, initially we developed an analytical framework of interference for NOMA in terms of signal to interference ratio (SIR). Later, we asses the outage probability of NOMA based downlink communication system by considering the effect of interference. The outage probability of NOMA with fixed number of interferers is calculated for different channel propagation effects as Nakagami-m, Rayleigh faded channels with and without log-normal shadowing. The obtained outage probabilities with fixed number of interferers are used to calculate the outage probabilities with random number of interferers (total system outage probability) in different channel propagation effects. In this paper, we proposed a novel algorithm to calculate the total system outage probability for NOMA in different channel propagation effects by choosing different offered load in terms of Erlangs per cell. We calculate the analytical results of outage probability for two users which are at near and edge positions of the cell. The obtained analytical results are supported with simulated NOMA.

It has been observed that Cloud services exhibit suboptimal performance for real-time requests due to increased network delay. Fog computing has emerged to address this issue by deploying Fog nodes at the network's edge. However, determining the optimal placement of Fog nodes for efficient service processing poses a significant challenge. The multiple ways to deploy a Fog node makes the Fog node placement an NP-class problem. It leverages the potential benefit of metaheuristic to solve this problem. In this work, we formulate a linear mathematical model for fog node placement (FNP), considering two important objectives for minimization, i.e., deployment cost (DC) and network latency (NL). A hybrid metaheuristic approach using genetic algorithm (GA) and JAYA, called JAYA-GA, is proposed to address this multi-objective optimization. The proposed model is simulated and the experimental results are compared against three popularly used metaheuristics: particle swarm optimization (PSO), GA, and JAYA. The proposed model consistently outperforms JAYA, PSO, and GA by the averages of 18.40%, 33.58%, and 30.75%, respectively, in terms of fitness (a weighted sum of DC and NL). Additionally, it exhibits superior performance on average convergence rate (16.60%, 53.42%, and 86.59%) and computation time (15.76%, 34.74%, and 59.22%) compared to JAYA, PSO, and GA respectively. Thus, the simulation results establish that the hybrid JAYA-GA technique surpasses the state-of-the-art alternatives on DC, NL, besides computation time and convergence rate.

Dental caries detection holds the key to unlocking brighter smiles and healthier lives by identifying one of the most common oral health issues early on. This vital topic sheds light on innovative ways to combat tooth decay, empowering individuals to take control of their oral health and maintain radiant smiles. This research paper delves into the realm of transfer learning techniques, aiming to elevate the precision and efficacy of dental caries diagnosis. Utilizing Keras ImageDataGenerator, a rich and balanced dataset is crafted by augmenting teeth images from the Kaggle teeth dataset. Five cutting-edge pre-trained architectures are harnessed in the transfer learning approach: EfficientNetV2B3, VGG19, InceptionResNetV2, Xception, and ResNet50, with each model, initialized using ImageNet weights and tailored top layers. A comprehensive set of evaluation metrics, encompassing accuracy, precision, recall, F1-score, and false negative rates are employed to gauge the performance of these architectures. The findings unveil the unique advantages and drawbacks of each model, illuminating the path to an optimal choice for dental caries detection using Grad-CAM (Gradient-weighted Class Activation Mapping). The testing accuracies achieved by EfficientNetV2B3, VGG19, InceptionResNetV2, Xception, and ResNet50 models stand at 95.89%, 96.58%, 93.15%, 93.15%, and 94.18%, respectively. The Training accuracies stood at 100%, 99.91%, 100%, 100% and 100%, meanwhile on validation we achieved 97.63%, 96.68%, 98.82%, 96.68%, and 100% accuracies for EfficientNetV2B3, VGG19, InceptionResNetV2, Xception, and ResNet50 models respectively. Capitalizing on transfer learning and juxtaposing diverse pre-trained architectures, this research paper paves the way for substantial advancements in dental diagnostic capabilities, culminating in enhanced patient outcomes and superior oral health.

Video description aims to translate the visual content in a video with appropriate natural language. Most of current works only focus on the description of factual content, paying insufficient attention to the emotions in the video. And the sentences always lack flexibility and vividness. In this work, a fact enhancement and emotion awakening based model is proposed to describe the video, making the sentence more attractive and colorful. The strategy of deep incremental leaning is employed to build a multi-layer sequential network firstly, and multi-stage training method is used to sufficiently optimize the model. Secondly, the modules of fact inspiration, fact reinforcement and emotion awakening are constructed layer by layer to discovery more facts and embed emotions naturally. The three modules are cumulatively trained to sufficiently mine the factual and emotional information. Two public datasets including EmVidCap-S and EmVidCap are employed to evaluate the proposed model. The experimental results show that the performance of the proposed model is superior to not only the baseline models, but also the other popular methods.

Deep reinforcement learning (DRL) is suitable for solving complex path-planning problems due to its excellent ability to make continuous decisions in a complex environment. However, the increase in the population size in the crowd evacuation path-planning problem causes a substantial computational burden for the algorithm, which leads to an unsatisfactory efficiency of the current DRL algorithm. This paper presents a path planning method based on DRL for crowd evacuation to solve the problem. First, we divide crowds into groups based on their relationship and distance from each other and select leaders from them. Next, we expand the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) to propose an Optimized Multi-Agent Deep Deterministic Policy Gradient (OMADDPG) algorithm to obtain the global evacuation path. The OMADDPG algorithm uses the Cross-Entropy Method (CEM) to optimize policy and improve the neural network’s training efficiency by applying the Data Pruning (DP) algorithm. In addition, the social force model is improved, incorporating the relationship between individuals and psychological factors into the model. Finally, this paper combines the improved social force model and the OMADDPG algorithm. The OMADDPG algorithm transmits the path information to the leaders. Pedestrians in the environment are driven by the improved social force model to follow the leaders to complete the evacuation simulation. The method can use a leader to guide pedestrians safely arrive the exit and reduce evacuation time in different environments. The simulation results prove the efficiency of the path planning method.