Digital microfluidic biochips (DMFBs), by precisely controlling and manipulating minute fluids, have realized the integration, automation, and cost-effectiveness of biochemical experiments and are applied in various fields such as medical diagnostics, drug development, and environmental monitoring. However, due to the composition of microelectronic component arrays, DMFBs are prone to electrode faults, leading to erroneous biochemical operations and, consequently, inaccurate experimental results. In this paper, a test path optimization algorithm combining an improved grey wolf algorithm and priority strategy is proposed to solve the problem of an extended test droplet path when DMFB tests faulty electrodes. By encoding the priority of the electrodes of the DMFB and the paths between the electrodes, the test droplet routing is performed according to the priority order. The priority coefficients are dynamically adjusted using the improved grey wolf algorithm to shorten the testing path length. Experimental results demonstrate that the proposed path optimization algorithm reduces the path length by 0.45% to 2.08% compared to the Eulerian circuit method in offline testing, achieving the theoretical optimum value, and by 4.90% to 10.53% compared to the ant colony algorithm in online testing. This provides an essential foundation for accelerating the safe and reliable development of DMFBs in healthcare.
The advent of the Internet has led to a heightened awareness of network security among users. While traditional intrusion detection models exhibit high detection accuracy, they require many labeled samples for model training. Furthermore, the emergence of unsupervised deep learning models has introduced the challenge of threshold selection. A novel flow intrusion detection system has been developed to address the challenges of data labeling and threshold selection. This system employs a Variational Auto-Encoder (VAE) model and a normalized MSE threshold selection method. It has been demonstrated to effectively identify flow intrusions, thereby reducing the burden of data labeling. The experimental results on the NSL-KDD dataset demonstrate that, compared with the traditional unsupervised algorithm, the VAE model exhibits a superior performance in terms of the area under the curve (AUC).
Ethernet Passive Optical Network (EPON) is a high-speed, energy-efficient fiber optic network that has gained attention with the growth of the Industrial Internet. However, existing security protocols for EPON, based on strand space theories, have flaws such as identity inconsistencies and consensus issues. This paper expands the strand space model and authentication testing method, enhances the Fiber Channel Password Authentication Protocol (FACP) strand space model in the EPON network environment using symmetric cryptography and asymmetric cryptography, fulfills the authentication requirements, and obtains the corresponding attack examples. Experimental results demonstrate that this protocol can enhance the security and efficiency of communication between devices and between devices and nodes (servers, disks) in commonly used EPON environments without affecting the identity authentication function.
Edge computing helps to release the tension at the center of IoT systems’ networks, thus reducing the latency, optimizing the bandwidth, and providing new privacy and security solutions, among others. Despite its benefits, edge computing faces unique challenges, including latency, security, and resource constraints. Among these challenges, energy consumption has emerged in the research community, and the global objective is to do "more with less". Researchers explore diverse strategies to enhance sustainability, from hardware optimizations to intelligent algorithms. The quest for energy efficiency and to reduce several impacts aligns with broader efforts to create an environmentally conscious technology landscape.In this paper, we present a task selection model for the edge. We focus on energy consumption and aim to maximize the value given by tasks, all the while minimizing the energy consumed. To do so, we develop a computer environment to simulate outdoor energy-harvesting edge devices, contribute to research reproducibility by recreating a photovoltaic energy harvesting prediction model, and train deep reinforcement learning models to select the best set of tasks at the edge. Our best deep reinforcement learning model, which uses Trust Region Policy Optimization, outperforms our best heuristic and is a robust task selector under varying external conditions.
Energy Harvesting (EH) from Radio Frequency (RF) energy sources has been envisioned as a promising approach to power small sensors, actuators and transceivers in the future Smart City. While several potential energy sources have been identified (mobile phones, dedicated power beacons, cellular base stations) most of them have tight power constraints or are too sparse to provide an adequate energy level. In the next decades, however, with the widespread of connected vehicles, vehicular communications will become a feasible source of RF energy. In this paper, we present the results of an experimental campaign aimed at quantifying the amount of energy that can be actually scavenged by battery-less sensors placed alongside a road from vehicular transmissions. Both a static and a dynamic scenario have been tested, in order to assess the impact of transmit power, vehicle trajectory and speed, as well as the storage capacity. In particular, our findings shed light on the peculiar features of this scenario, which may not be fully leveraged by commonly used harvesting boards tailored to different types of energy sources. An upper bound of the available energy for the dynamic scenario is derived, confirming that up to 0.45 mJ can be stored from the passage of a single vehicle with a proper tuning of the parameters setup.
With the continuous advancement of blockchain technology, smart contracts, as one of its core applications, have increasingly become a focal point for security concerns. To address this, this paper proposes a novel method that integrates keyword filter technology, the pre-trained Bidirectional Encoder Representations from Transformers (BERT) model, and the Bidirectional Long Short-Term Memory-Conditional Random Field (BiLSTM-CRF) architecture, aimed at enhancing the efficiency and accuracy of vulnerability detection in Solidity smart contracts. Initially, keyword filter technology is employed to select and preprocess code snippets, extracting features closely associated with security vulnerabilities. Subsequently, the BERT model conducts deep semantic analysis and feature extraction, after which the BiLSTM-CRF architecture further learns from the features and predicts vulnerability types. Extensive experiments conducted on a dataset comprising eight major types of vulnerabilities demonstrate that the method proposed in this study significantly outperforms existing vulnerability detection methods in key metrics such as accuracy, recall, and F1 scores. This research not only provides an effective technical solution for detecting security vulnerabilities in smart contracts but also holds significant theoretical and practical implications for promoting the safe and reliable development of blockchain technology.
With each advancement in internet technology, new security challenges arise. The prevalence of malicious programs continues to increase, which makes it crucial to detect and address them effectively. Many researchers focus on solving different datasets by using deep learning methods and make significant progress. However, these strategies must be continuously improved to adapt to the latest data. In this paper, an improved model based on CNN-LSTM is proposed to detect and classify malware programs, named malDetect I. At the same time, the Transformer Encoder module is also modified based on model Bert to adapt to the classification task. Lastly, two models are compared with prediction results on evaluation indicators. The data used in this paper is the Windows API sequence extracted after dynamic operation. The text processing methods are also suitable for processing sequence data. The experiment uses Word2Vec and two different learning rate strategies, and the improved model accuracy is 9.83% higher than the original CNN-LSTM model. The model integrated the BiLSTM model with the Self-Attention mechanism, named malDetect II, is 11.46% higher than the basic model CNN-LSTM and 2.82% higher than the Transformer Encoder classification model.
In Internet of Things (IoT) anomaly detection (AD) tasks, reconstruction-based AD methods are widely concerned due to their simple network structure and ease of training. However, they tend to directly use the original image as the output due to their powerful generalization ability. To overcome this, we propose a novel reconstruction-based anomaly detection method based on local key features and generative adversarial network (named Local-GAN) for IoT. Firstly, Local-GAN utilizes the dual feature encoders to extract local features, and select key features with masks generated by a mask generator. Based on the selected local key features, the decoder avoids reconstructing the anomaly sample/area as it is, which improves the performance. Secondly, the quality-enhanced discriminator is imported into the model to improve the quality of reconstructed images, making it easier to distinguish anomalies. The experiment results on 3 image datasets and 1 real-world industrial image dataset show that our method outperforms reconstruction-based anomaly detection methods in image AD tasks. Furthermore, the experiment results on 13 structured datasets demonstrate the generality of our method.
With the continuous development of Machine Learning as a Service (MLaaS), model stealing has become an emerging problem in machine learning security in recent years. In model stealing, one typically obtains the soft labels of model queries and a proxy dataset as prior knowledge, but this scenario is highly idealised. How to steal models without data and hard labels is a pressing problem that needs to be solved. The current mainstream of model stealing attack methods mainly focus on stealing the accuracy of the model and overlook the robustness of the model. However, robustness is essential in security applications such as facial recognition and secure payment scenarios. Moreover, building robust models usually requires costly adversarial training and fine-tuning, making these models the primary targets for theft. To address these issues, in this paper, we propose a new data-free robustness stealing method under data-free conditions from the perspective of data generation, thereby better shaping the classification boundary data to optimise the accuracy and robustness of the models. Through testing, our method achieved clean accuracy and robust accuracy of 53.69% and 21.0%, respectively, under the more complex CIFAR-100 dataset classification. These results are only 3.06% and 3.94% different from the target model, respectively, showing a significant improvement over recent research.
The efficient management of power consumption on operators’ networks is becoming crucial for network economics, reliability, and sustainability, as longs as the energy consumption of network devices increases with network usage and growth in scale. Nowadays, there are no means of automatically retrieving from the network timely information about the actual energy consumption on different communications paths. This paper concentrates on the creation of topological maps for IP networks (from access to backbone) capturing the current power consumption across paths, by leveraging on standard solutions for that purpose. The resulting maps can be then leveraged as an input for smart and informed decisions for optimization purposes.
Nowadays, swarm intelligence algorithms are used to solve various problem in IoT environments because of their excellent performance, and the particle swarm algorithm(PSO) is a superior algorithm in swarm intelligence algorithms. The original PSO, however, exhibits certain limitations. For instance, the utilization of random initialization results in a less diverse population distribution for PSO. During the algorithm iteration process, excessively large or small weight factors can also detrimentally impact the performance of the algorithm. Moreover, when dealing with multi-modal problems, the PSO algorithm can quickly converge to local optimal because of its "precocious". In this paper, a Double-Population Particle Swarm Optimization(DP-PSO) algorithm is proposed. Firstly, the chaotic mechanism is used to initialize the population to make its distribution more uniform and obtain better convergence results. Secondly, the oscillation characteristic of the sine function is used to improve the updating method of weight by adding the sine factor, and the search area of the particle is expanded. Finally, the double-population variation mechanism is introduced to enhance the variation ability and improve the solving accuracy in the later iteration. The experimental results show that the DP-PSO is superior to other algorithms in performance and has obvious improvement in the ability to expand the search space of the population and jump out of the local optimal solution. When applied to cantilever beam design, DP-PSO also has good competitiveness in terms of stability, robustness, and scalability.
The Internet of Things is rapidly advancing, enabling interactions with the physical world and presenting various applications. Routing Protocol over Low-Power and Lossy Networks (RPL) is the routing protocol introduced by the IETF for IoT environments. RPL has been subjected to multiple attacks, making it essential to develop a model that defines its behavior. To achieve this, we propose DEVS-RPL, which employs the discrete event model to specify the normal behavior of network nodes in generating RPL messages. DEVS-RPL estimates the mean and frequency of each generated RPL message by defining the participant states and utilizing the time advance function. As a proof of concept, the normal network traffic produced by the Cooja simulator is compared to the RPL messages generated by the DEVS-RPL model. We find that the mean absolute percentage error (MAPE) of DEVS-RPL is less than 9%, indicating the high accuracy of the proposed model.
The advancement of Natural Language Processing (NLP) technology has led to a focus on language models (LMs) with a general-purpose, which have achieved greater accuracy than earlier methods. However, specialized LMs still face challenges, particularly in domains with limited resources. To address this issue, this paper proposes a Multi-Domain and Multi-View oriented Deep Neural Network (MDMV-DNN) model for sentiment analysis in Large Language Models (LLMs). A novel architecture, which incorporates a Domain-Integrated Residual Attention enhanced BERT (DIRA-BERT) mechanism for multi-domain integration, and a Multi-Filter Convolutional Neural Network (MF-CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) enhanced fusion mechanism for Multi-View feature extraction, is introduced and constructed. The DIRA-BERT can enhance the classification accuracy by integrating multiple expert BERT models tailored to specific tasks, thus providing refined and interpretable feature representations. The integration of MF-CNN and Bi-LSTM can further strengthen the learning ability in capturing the local key and sequential features respectively, so as to enable a comprehensive sentiment analysis through multi-view feature extraction. Experiment results demonstrate that the proposed MDMV-DNN consistently outperforms the traditional methods using common BERT models across all evaluation metrics, showcasing its effectiveness in specialized sentiment analysis tasks.
Large Language Models (LLMs) have gained popularity since the release of ChatGPT in 2022. These systems utilize Artificial Intelligence (AI) algorithms to analyze natural language, enabling users to have sophisticated real-time conversations with them. The existing literature on LLMs is mostly focused on system design and lacks dedicated research on investigating privacy and security issues. To safeguard the interests of various stakeholders, it is crucial to understand the associated security and privacy risks of these models. Our study utilized STRIDE and LINDDUN methodologies to investigate security and privacy threats of LLMs. We presented a detailed system model of LLMs and analyzed the potential threats, vulnerabilities, security considerations, and mitigation tactics intrinsic to the design and deployment of various system components. Our comprehensive threat assessment showcases potential threats imminent to the current generation of LLMs, such as unintentional data leakage or system misuse by malicious actors. Furthermore, our study discusses the importance of proactive security measures in LLM development, deployment, and maintenance.
IoT devices offer an efficient solution for collecting visual data for inspections, including identifying potential cracks in infrastructure. Drones can capture precise images or videos, which are then transmitted to the cloud for processing or analyzed using fog computing nodes. Edge computing allows on-site data analysis, optimizing performance and reducing latency. Recently, Visual Crack Detection (VCD) using low-cost cameras has gained attention in Structural Health Monitoring (SHM) for maintenance and damage identification. Crack Detection (CD) algorithms face challenges from brightness changes, shadows, and complex backgrounds. Deep Learning (DL) networks have been developed to improve crack identification, but most use local attributes and pixel correlations. A proposed multi-scale convolutional network inspired by Self-Attention (SA) combines local and global attributes for more robust crack detection. The performance of this multi-scale convolutional network has been tested on some well-known crack datasets and evaluated against state-of-the-art approaches. The test results on the well-known Crack500 dataset outperforms the existing approaches.
Nowadays, the Internet of Things (IoT) is reshaping numerous application domains. Amidst the different communication protocols currently available (MQTT, CoAP, AMQP, DPWS, etc.) and the abundance of management platforms, the IoT domain has stumbled into vertical silos of proprietary systems hindering interoperability. Within this intricate and fragmented ecosystem, the need for an intelligent and scalable architecture promoting interoperability becomes imperative to maximize the potential of IoT. This paper introduces a system designed to enhance the convergence of IoT protocols (e.g., MQTT, CoAP, AMQP, etc.) with the Web of Things (WoT) paradigm. Our proposed gateway-based solution integrates two systems: Stack4Things (S4T) and Data eXchange Mediator Synthesizer (DeXMS). The role of the latter is to adapt, at the gateway level, the IoT protocols and expose the resources/functionalities of the IoT devices as RESTful APIs using HTTP. Meanwhile, the former (i.e., S4T), leveraging its Dynamic DNS system, ensures that these RESTful resources (exposed at the gateway) are accessible over the Web using publicly routable Uniform Resource Locators (URLs) even when the gateway is deployed behind networking middleboxes (e.g., NATs and firewalls).
Green supercomputing has been attracting more and more attention due to the acknowledgement of environmental issues. This is especially the case in the field of supercomputing: given the large number of compute nodes involved, the power consumption of these machines is high, in the order of tens of megawatts. Efficient usage of supercomputers is thus key to minimise their power consumption, and one approach is to increase data transfer rates so that the computer can be returned into an energy efficient sleep mode rapidly. Parallel paths and especially disjoint paths routing is one solution to achieve such more efficient data transfers. Yet, the selection of mutually disjoint paths is complex, and in some cases graph enumeration can produce a solution whereas conventional approaches fail. The contribution of this paper is twofold: for the selected three fundamental graph enumeration problems, we give simple, easy to understand proofs that do not involve complex mathematical notions as in previous works; we describe and evaluate the corresponding graph generation algorithms and make comparisons to straightforward or conventional approaches. Not only theoretical evaluation but also empirical evaluation is conducted. Time complexity is shown to be optimal.
In recent years, federated learning has been widely used in deep learning training tasks such as image recognition. The federated learning mechanism allows multiple participants to collaboratively train a common model without having to aggregate data. However, participants may infringe on the privacy of data owners when collecting personal data. And when there are significant differences in data distribution among participants, traditional federated learning methods may yield less than ideal training results. To solve these problems, we propose an efficient privacy-preserving collaborative learning framework (EPPCL), which provides secure access authorization and training for personal data. While introducing an authorization mechanism, we design an efficient collaborative training method that can improve the model performance in cases of data heterogeneity. Furthermore, we evaluate the system performance of EPPCL. The experimental results demonstrate that EPPCL is a secure and effective way to achieve multi-party collaborative training for personal data.
The rapid growth of Internet of Things (IoT) results in tremendous interest in research for low-power wireless sensors. The energy harvesting (EH) ability of IoT sensors determines the efficiency and reliability of network connectivity inside IoT. This work considers a wireless sensor network (WSN) where a fusion center (FC) gathers data packets from EH sensors which can store energy without battery overflow or leakage. The FC selects a subset of nodes per time slot to gather data from them under stochastic data arrival processes via its mutually orthogonal channels. EH processes and battery states of sensors are unknown to FC. Similarly, data arrival (DA) processes and buffer states of sensors are unknown to FC, which only knows previous transmission results. We aim to propose a simple, efficient policy that maximizes network throughput and fairness in IoT network systems, which is very important, especially for 5G and next-generation networks. Data transmission relies on not only scheduled nodes’ gathered energy but also their buffered data. If it has data to send and sufficient energy for transmission, a node can transmit data when scheduled. This paper proposes a low-complexity algorithm that is almost throughput and fairness optimal for quite general EH and DA processes over finite time horizons. We show that by removing the battery capacity limit, the proposed approach yields asymptotically optimal throughput and fairness for general EH and DA processes over an infinite time horizon. It achieves nearly optimality in throughput and fairness across finite time horizons with finite-capacity batteries, according to numerical simulations whereas existing solutions become suboptimal.
In the near future, drones will play a key role in streamlining the delivery processes between businesses and their clients by utilizing high-accuracy localization. This study introduces a system that employs a 5G network, enhanced by multiple radio access technologies (RAT), to address the challenges of accurate positioning for drones. For indoor settings, where the 5G signal strength is not as robust as it is outdoors, we use Wireless Fidelity (Wi-Fi) routers, here referred to as Wi-5G TRPs, to enhance location accuracy. These points connect through the 5G infrastructure to relay information to the same location management function (LMF) that manages the cellular 5G network. The core of our research demonstrates how the placement of Wi-5G TRPs inside buildings significantly influences how accurately they can pinpoint a location. By developing an innovative deep reinforcement learning (DRL) framework, we tackle the complex challenge of determining the most effective arrangement of Wi-5G TRPs to achieve high-accuracy three-dimensional (3D) tracking of moving targets. Our findings indicate that the overall location accuracy is influenced by both the measurement errors and the spatial error distribution, known as geometric dilution of precision (GDOP). Our extensive tests confirm that inaccuracies are predominantly due to vertical measurements rather than horizontal ones. Lastly, by analyzing the outcomes of strategically placing the Wi-5G TRPs utilizing our proposed DRL-based solution, we can reduce the location estimation errors compared to arbitrary placements.
