Advanced Persistent Threats (APTs) have significantly impacted organizations over an extended period with their coordinated and sophisticated cyberattacks. Unlike signature-based tools such as antivirus and firewalls that can detect and block other types of malware, APTs exploit zero-day vulnerabilities to generate new variants of undetectable malware. Additionally, APT adversaries engage in complex relationships and interactions within network entities, necessitating the learning of interactions in network traffic flows, such as hosts, users, or IP addresses, for effective detection. However, traditional deep neural networks often fail to capture the inherent graph structure and overlook crucial contextual information in network traffic flows. To address these issues, this research models APTs as heterogeneous graphs, capturing the diverse features and complex interactions in network flows. Consequently, a hetero-geneous graph transformer (HGT) model is used to accurately distinguish between benign and malicious network connections. Experiment results reveal that the HGT model achieves better performance, with 100 \% accuracy and accelerated learning time, outperferming homogeneous graph neural network models.

Low probability of detection (LPD) has recently emerged as a means to enhance the privacy and security of wireless networks. Unlike existing wireless security techniques, LPD measures aim to conceal the entire existence of wireless communication instead of safeguarding the information transmitted from users. Motivated by LPD communication, in this paper, we study a privacy-preserving and distributed framework based on graph neural networks to minimise the detectability of a wireless ad-hoc network as a whole and predict an optimal communication region for each node in the wireless network, allowing them to communicate while remaining undetected from external actors. We also demonstrate the effectiveness of the proposed method in terms of two performance measures, i.e., mean absolute error and median absolute error.

Vehicular Ad Hoc Networks (VANETs) have the capability of swapping every node of every individual while driving and traveling on the roadside. The VANET-connected vehicle can send and receive data such as requests for emergency assistance, current traffic conditions, etc. VANET assistance with a vehicle for communication purposes is desperately needed. The routing method has the characteristics of safe routing to repair the trust-based features on a specific node.When malicious activity is uncovered, intrusion detection systems (IDS) are crucial tools for mitigating the damage. Collaborations between vehicles in a VANET enhance detection precision by spreading information about interactions across their nodes. This makes the machine learning distribution system feasible, scalable, and usable for creating VANET-based cooperative detection techniques. Privacy considerations are a major impediment to collaborative learning due to the data flow between nodes. A malicious node can get private details about other nodes by observing them. This study proposes a cooperative IDS for VANETs that safeguards the data generated by machine learning. In the intrusion detection phase, the selected optimal characteristics is used to detect network intrusion via a hybrid Deep Neural Network and Bidirectional Long Short-Term Memory approach. The Trust-based routing protocol then performs the intrusion prevention process, stopping the hostile node by having it select the most efficient routing path possible.

This paper investigates the output feedback security control problem of switched nonlinear systems (SNSs) against denial-of-service (DoS) attacks. A novel switched observer-based neural network (NN) adaptive control algorithm is established, which guarantees that all the signals in the closed-loop system remain bounded. Note that when a DoS attacker is active in the Sensor-Controller channel, the controller cannot acquire accurate information, which leads to the standard backstepping technique not being workable. A set of NN adaptive switching-like observers is designed to tackle the obstacle for each subsystem. Further, by combining the proposed observer with the backstepping technique, an NN adaptive controller is constructed and the dynamic surface control method is borrowed to surmount the complexity explosion phenomenon. Finally, an illustrative example is provided to demonstrate the effectiveness of the proposed control algorithm.

In the past two years, technology has undergone significant changes that have had a major impact on healthcare systems. Artificial intelligence (AI) is a key component of this change, and it can assist doctors with various healthcare systems and intelligent health systems. AI is crucial in diagnosing common diseases, developing new medications, and analyzing patient information from electronic health records. However, one of the main issues with adopting AI in healthcare is the lack of transparency, as doctors must interpret the output of the AI. Explainable AI (XAI) is extremely important for the healthcare sector and comes into play in this regard. With XAI, doctors, patients, and other stakeholders can more easily examine a decision s reliability by knowing its reasoning due to XAI s interpretable explanations. Deep learning is used in this study to discuss explainable artificial intelligence (XAI) in medical image analysis. The primary goal of this paper is to provide a generic six-category XAI architecture for classifying DL-based medical image analysis and interpretability methods.The interpretability method/XAI approach for medical image analysis is often categorized based on the explanation and technical method. In XAI approaches, the explanation method is further sub-categorized into three types: text-based, visualbased, and examples-based. In interpretability technical method, it was divided into nine categories. Finally, the paper discusses the advantages, disadvantages, and limitations of each neural network-based interpretability method for medical imaging analysis.

This article presents a new concept of fully analogue adaptive filters. The adaptation is based on fully analogue neural networks. With the use of a filter bank, it can be used for high frequency and real-time adaptation. The properties of this concept are verified using electronic circuit simulations.

Frequency hopping (FH) technology is one of the most effective technologies in the field of radio countermeasures, meanwhile, the recognition of FH signal has become a research hotspot. FH signal is a typical non-stationary signal whose frequency varies nonlinearly with time and the time-frequency analysis technique provides a very effective method for processing this kind of signal. With the renaissance of deep learning, methods based on time-frequency analysis and deep learning are widely studied. Although these methods have achieved good results, the recognition accuracy still needs to be improved. Through the observation of the datasets, we found that there are still difficult samples that are difficult to identify. Through further analysis, we propose a horizontal spatial attention (HSA) block, which can generate spatial weight vector according to the signal distribution, and then readjust the feature map. The HSA block is a plug-and-play module that can be integrated into common convolutional neural network (CNN) to further improve their performance and these networks with HSA block are collectively called HANets. The HSA block also has the advantages of high recognition accuracy (especially under low SNRs), easy to implant, and almost no influence on the number of parameters. We verified our method on two datasets and a series of comparative experiments show that the proposed method achieves good results on FH datasets.

With the increased commercialization of deep learning (DL) models, there is also a growing need to protect them from illicit usage. For cost- and ease of deployment reasons it is becoming increasingly common to run DL models on the hardware of third parties. Although there are some hardware mechanisms, such as Trusted Execution Environments (TEE), to protect sensitive data, their availability is still limited and not well suited to resource demanding tasks, like DL models, that benefit from hardware accelerators. In this work, we make model stealing more difficult, presenting a novel way to divide up a DL model, with the main part on normal infrastructure and a small part in a remote TEE, and train it using adversarial techniques. In initial experiments on image classification models for the Fashion MNIST and CIFAR 10 datasets, we observed that this obfuscation protection makes it significantly more difficult for an adversary to leverage the exposed model components.

Advanced persistent threat (APT) attacks have caused severe damage to many core information infrastructures. To tackle this issue, the graph-based methods have been proposed due to their ability for learning complex interaction patterns of network entities with discrete graph snapshots. However, such methods are challenged by the computer networking model characterized by a natural continuous-time dynamic heterogeneous graph. In this paper, we propose a heterogeneous graph neural network based APT detection method in smart grid clouds. Our model is an encoderdecoder structure. The encoder uses heterogeneous temporal memory and attention embedding modules to capture contextual information of interactions of network entities from the time and spatial dimensions respectively. We implement a prototype and conduct extensive experiments on real-world cyber-security datasets with more than 10 million records. Experimental results show that our method can achieve superior detection performance than state-of-the-art methods.

Propagation delay and channel loss are two vital factors affecting reliability of Underwater Acoustic Networks (UANs). Different from land networks, UANs have long propagation delay and poor channel quality, which lead to serious data collision and high bit error rate, respectively. However, complex underwater environments impose great challenges to evaluate propagation delay and channel loss. As temperature is the most critical factor affecting them, in this paper, we propose to employ temperature to evaluate them. However, existing temperature prediction research are insufficient for accuracy or efficiency. This paper proposes a temperature prediction-assisted approach for evaluating propagation delay and channel loss, aiming to improve reliability and performance of underwater acoustic networks. We build a nonlinear autoregressive dynamic neural network-based temperature prediction model to improve prediction accuracy and reduce time complexity. Then, we evaluate propagation delay and channel loss considering different marine environments, including shallow and deep sea. Extensive simulation results show that our approach performs better than five advanced baselines.

With the rapid development of underwater sensor networks, the design of underwater demodulators become increasingly significant. However, underwater acoustic communication is faced with many problems such as propagation time delay, multipath effect and Doppler effect due to the complexity of underwater environment. Demodulation of underwater communication signals is a challenging task. To solve this problem, we propose a novel binary phase shift keying (BPSK) demodulator for underwater acoustic communication based on convolutional neural network, which demodulates the modulation data by detecting the position of phase shift. The method proposed in this paper significantly reduces the bit error rate (BER) compared with the results of the traditional method in URPC1 datasets (Underwater Robot Picking Contest).

Understanding dynamic human behavior based on online video has many applications in security control, crime surveillance, sports, and industrial IoT systems. This paper solves the problem of classifying video data recorded on surveillance cameras in order to identify fragments with instances of shoplifting. It is proposed to use a classifier that is a symbiosis of two neural networks: convolutional and recurrent. The convolutional neural network is used for extraction of features from each frame of the video fragment, and the recurrent network for processing the temporal sequence of processed frames and subsequent classification.

Wearables Security 2022 - Mobile devices such as smartphones are increasingly being used to record personal, delicate, and security information such as images, emails, and payment information due to the growth of wearable computing. It is becoming more vital to employ smartphone sensor-based identification to safeguard this kind of information from unwanted parties. In this study, we propose a sensor-based user identification approach based on individual walking patterns and use the sensors that are pervasively embedded into smartphones to accomplish this. Individuals were identified using a convolutional neural network (CNN). Four data augmentation methods were utilized to produce synthetically more data. These approaches included jittering, scaling, and time-warping. We evaluate the proposed identification model’s accuracy, precision, recall, F1-score, FAR, and FRR utilizing a publicly accessible dataset named the UIWADS dataset. As shown by the experiment findings, the CNN with the timewarping approach operates with very high accuracy in user identification, with the lowest false positive rate of 8.80\% and the most incredible accuracy of 92.7\%.

Wearables Security 2022 - One of the biggest new trends in artificial intelligence is the ability to recognise people s movements and take their actions into account. It can be used in a variety of ways, including for surveillance, security, human-computer interaction, and content-based video retrieval. There have been a number of researchers that have presented vision-based techniques to human activity recognition. Several challenges need to be addressed in the creation of a vision-based human activity recognition system, including illumination variations in human activity recognition, interclass similarity between scenes, the environment and recording setting, and temporal variation. To overcome the above mentioned problem, by capturing or sensing human actions with help of wearable sensors, wearable devices, or IoT devices. Sensor data, particularly one-dimensional time series data, are used in the work of human activity recognition. Using 1D-Convolutional Neural Network (CNN) models, this works aims to propose a new approach for identifying human activities. The Wireless Sensor Data Mining (WISDM) dataset is utilised to train and test the 1D-CNN model in this dissertation. The proposed HAR-CNN model has a 95.2\%of accuracy, which is far higher than that of conventional methods.

Provenance 2022 - Advanced Persistent Threats (APTs) are typically sophisticated, stealthy and long-term attacks that are difficult to be detected and investigated. Recently proposed provenance graph based on system audit logs has become an important approach for APT detection and investigation. However, existing provenance-based approaches that either require rules based on expert knowledge or cannot pinpoint attack events in a provenance graph still cannot effectively mitigate APT attacks. In this paper, we present Deepro, a provenance-based APT campaign detection approach that not only effectively detects attack-relevant entities in a provenance graph but also precisely recovers APT campaigns based on the detected entities. Specifically, Deepro first customizes a general purpose GNN (Graph Neural Network) model to represent and detect process nodes in a provenance graph through automatically learning different patterns of attack behaviors and benign behaviors using the rich contextual information in the provenance graph. Then, Deepro further detects attack-relevant file and network entities according to their data dependencies with the detected process nodes. Finally, Deepro recovers APT campaigns through correlating detected entities based on their causality relationships in the provenance graph. We evaluated Deepro with ten real-world APT attacks. The evaluation result shows that Deepro can effectively detect attack events with an average 98.81\% F1-score and thus produces precise provenance sub-graphs of APT attacks.

Neural Style Transfer - With the development of economical society, the problem of product piracy security is becoming more and more serious. In order to protect the copyright of brands, based on the image neural style transfer, this paper proposes an automatic generation algorithm of anti-counterfeiting logo with security shading, which increases the difficulty of illegal copying and packaging production. VGG19 deep neural network is used to extract image features and calculate content response loss and style response loss. Based on the original neural style transfer algorithm, the content loss is added, and the generated security shading is fused with the original binary logo image to generate the anti-counterfeiting logo image with higher recognition rate. In this paper, the global loss function is composed of content loss, content response loss and style response loss. The L-BFGS optimization algorithm is used to iteratively reduce the global loss function, and the relationship between the weight adjustment, the number of iterations and the generated anti-counterfeiting logo among the three losses is studied. The secret keeping of shading style image used in this method increases the anti-attack ability of the algorithm. The experimental results show that, compared with the original logo, this method can generate the distinguishable logo content, complex security shading, and has convergence and withstand the attacks.

Neural Style Transfer - As one of the fields of computer art creation, style transfer has become more and more popular. However, in order to obtain good visual effects, a large number of neural style transfer algorithms use semantic map to guide the style transfer between the correct regions. As an important means to ensure the quality of style transfer, semantic map can meaningfully control the results of style transfer. However, the method of manually generating semantic graph is cumbersome and inefficient. In this paper, we introduce a semantic segmentation network to automatically generate the semantic map required by neural style transfer, and combine it with neural style transfer network, we propose a new neural style transfer algorithm. Experiments show that our algorithm not only avoids cumbersome manual work, but also generates high-quality style transfer results.

Neural Style Transfer - Image style transfer is an important research content related to image processing in computer vision. Compared with traditional artificial computing methods, deep learning-based convolutional neural networks in the field of machine learning have powerful advantages. This new method has high computational efficiency and a good style transfer effect. To further improve the quality and efficiency of image style transfer, the pre-trained VGG-16 neural network model and VGG-19 neural network model are used to achieve image style transfer, and the transferred images generated by the two neural networks are compared. The research results show that the use of the VGG-16 convolutional neural network to achieve image style transfer is better and more efficient.

Neural Network Security - Trust is an essential concept in ad hoc network security. Creating and maintaining trusted relationships between nodes is a challenging task. This paper proposes a decentralized method for evaluating trust in ad hoc networks. The method uses neural networks and local information to predict the trust of neighboring nodes. The method was compared with the original centralized version, showing that even without global information knowledge, the method has, on average, 97\% accuracy in classification and 94\% in regression problem. An important contribution of this paper is overcoming the main limitation of the original method, which is the centralized evaluation of trust. Moreover, the decentralized method output is a perfect fit to use as an input to enhance routing in ad hoc networks.

Neural Network Security - Software-Defined Network (SDN) is a new networking paradigm that adopts centralized control logic and provides more control to the network operators over the network infrastructure to meet future network requirements. SDN controller known as operation system, which is responsible for running network applications and maintaining the different network services and functionalities. Despite all its great capabilities, SDN is facing different security threats due to its various architectural entities and centralized nature. Distributed Denial of Service (DDoS) is a promptly growing attack and becomes a major threat for the SDN. To date, most of the studies focus on detecting high-rate DDoS attacks at the control layer of SDN and low-rate DDoS attacks are high concealed because they are difficult to detect. Furthermore, the existing methods are useful for the detection of high-rate DDoS, so need to focus on low-rate DDoS attacks separately. Hence, the use of machine learning algorithms is growing for the detection of low-rate DDoS attacks in the SDN, but they achieved low accuracy against this attack. To improve the detection accuracy, this paper first describes the attack s mechanism and then proposes a Recurrent Neural Network (RNN) based method. The extracted features from the flow rules are used by the RNN for the detection of low-rate attacks. The experimental results show that the proposed method intelligently detects the attack, and its detection accuracy reaches 98.59\%. The proposed method achieves good detection accuracy as compared to existing studies.

Neural Network Security - Aiming at the network security problem caused by the rapid development of network, this paper uses a network traffic anomaly detection method of industrial control system based on convolutional neural network. In the traditional machine learning algorithm, the processing of features has a high impact on the performance of the model, and the model is highly dependent on features. This method uses the characteristics of convolutional neural network to autonomously learn features, which avoids this problem. In order to verify the superiority of the model, this paper takes accuracy as the evaluation index, and compares it with the traditional machine learning algorithm. The results show that the overall accuracy of the method is 99.88 \%, which has higher accuracy than traditional machine learning algorithms such as decision tree algorithm (ID3), adaptive boosting tree (Adboost) and naive Bayesian model. Therefore, this method can be better applied to the anomaly detection of network traffic in industrial control system, and has practical application value.

Neural Network Security - With the development of computer and network technology, industrial control systems are connecting with the Internet and other public networks in various ways, viruses, trojans and other threats are spreading to industrial control systems, industrial control system information security issues are becoming increasingly prominent. Under this background, it is necessary to construct the network security evaluation model of industrial control system based on the safety evaluation criteria and methods, and complete the safety evaluation of the industrial control system network according to the design scheme. Based on back propagation (BP) neural network’s evaluation of the network security status of industrial control system, this paper determines the number of neurons in BP neural network input layer, hidden layer and output layer by analyzing the actual demand, empirical equation calculation and experimental comparison, and designs the network security evaluation index system of industrial control system according to factors affecting industrial control safety, and constructs a safety rating table. Finally, by comparing the performance of BP neural network and multilinear regression to the evaluation of the network security status of industrial control system through experimental simulation, it can be found that BP neural network has higher accuracy for the evaluation of network security status of industrial control system.

Neural Network Security - With the continuous development of network technology and the continuous expansion of network scale, the security of the network has suffered more threats, and the attacks faced are becoming more and more extensive. The frequent occurrence of network security incidents has caused huge losses, facing more and more severe situation, it is necessary to adopt various network security technologies to solve the problem. In network security, the most commonly used technology is firewall. The firewall has a certain blocking effect on attacks from outside the network, but it has a weak defense effect on the attacks in the internal network, and it is easy to be bypassed. Intrusion detection technology can detect both internal and external network attacks. Responses are generated before the intrusion behavior occurs, and alarm information is issued for timely and effective processing. In recent years, China s campus security incidents are still happening, seriously threatening the lives of students and disrupting the normal teaching order of schools. At present, there are still many loopholes in campus security operations. Campus security management system has become an important task in campus security construction. On this basis, relevant personnel are required to analyze the existing problems of campus safety and the needs of the safety management system, and find the main technology of a more advanced intelligent safety management system.

Neural Network Security - With the rapid development of computer networks and information technology today, people are more inclined to use network systems to achieve various data exchanges. Alibaba, Tencent and other companies virtual payment has become the mainstream payment method. Due to the globalization and openness of the network, anyone can freely enter and exit, which brings huge hidden dangers to NS(network security). NS has become an important issue that we have to face. Once important information is stolen, it is likely to cause very large losses to individuals and even the society. This article mainly studies the computer NS encryption technology of neural network. First of all, the current situation of computer NS is comprehensively reflected from the two aspects of domestic Internet users and NS penetration rate in recent years. By 2020, the number of Chinese residents using the Internet has reached 1.034 billion, and 77.3\% of Internet users are generally aware of NS. Secondly, it analyzes the effect of NN(neural network) on computer NS encryption technology. The results show that the use of NN in computer encryption technology not only helps to improve security and convenience, but also prevents the secondary transmission of data and prevents related information leakage.

Neural Network Security - With the development of computing technology, data security and privacy protection have also become the focus of researchers; along with this comes the issue of network link security and reliability, and these issues have become the focus of discussion when studying network security. Intrusion detection is an effective means to assist in network malicious traffic detection and maintain network stability; to meet the ever-changing demand for network traffic identification, intrusion detection models have undergone a transformation from traditional intrusion detection models to machine learning intrusion detection models to deep intrusion detection models. The efficiency and superiority of deep learning have been proven in fields such as image processing, but there are still some problems in the field of network security intrusion detection: the models are not targeted when processing data, the models have poor generalization ability, etc. The combinatorial neural network proposed in this paper can effectively propose a solution to the problems of existing models, and the CL-IDS model proposed in this paper has a better performance on the KDDCUP99 dataset as demonstrated by relevant experiments.