ArchiveWe explore the transition from closed environments to open world environments and its impact on visual perception (focusing on object recognition and detection) and the field of deep learning. In open world environments, software systems need to adapt to constantly changing conditions and demands, presenting new challenges for deep learning methods. In particular, open world visual perception requires systems to understand and process environments and objects not seen during the training phase, which exceeds the capabilities of traditional closed systems. We first discuss the dynamic and adaptive system requirements brought about by technological advances, highlighting the advantages of open systems over closed systems. Then we delve into the definition of the open world and existing work, covering five dimensions of openness: open set learning, zero-shot learning, few-shot learning, long-tail learning, and incremental learning. In terms of open world recognition, we analyze the core challenges of each dimension and provide quantified evaluation metrics for each task dataset. For open world object detection, we discuss additional challenges compared with recognition, such as occlusion, scale, posture, symbiotic relationships, background interference, etc., and emphasize the importance of simulation environments in constructing open world object detection datasets. Finally, we underscore the new perspectives and opportunities that the concept of the open world brings to deep learning, acting as a catalyst for technological advancement and deeper understanding of the realistic environment challenges, offering a reference for future research.
Autonomous driving, connected vehicles, electrification of the powertrain, and shared mobility lead to the rapid increasing of the complexity of automotive electronic system, and the functional safety and cyber-security problems of the automotive electronic system cause a serial of recalls frequently, which is resulting in a huge economic loss and user experience decline of the original equipment manufacturer (OEM). The over-the-air (OTA) update technology uses wireless network to achieve remote update of software and firmware in scenarios such as automatic driving function update, on-board software update and on-board safety system upgrade, thus avoiding the adverse impact of recall, but how to guarantee the cyber-security and efficient implementation of the OTA processes is a key problem needing to be resolved by auto industry. The opensource Uptane framework is an industry reference specification for OTA, but it has some cyber-security vulnerabilities and brings too large system overhead for its current reference implementation. By choosing efficient Hash and signature algorithms and introducing a new verification mechanism based on alliance chain, an optimized Uptane framework is proposed with reduced resource and time overhead and increased cyber-security. The prototype implementation verifies the cyber-security of the optimized Uptane framework, and by comparing with the original Uptane framework, the memory consumption and delay overhead of the optimized Uptane framework are reduced by 6.9% and 28.6%, respectively.
Existing electric vehicle API platforms (e.g., SmartCar), use access control mechanisms to preserve users’ privacy. To preserve location privacy and meanwhile enable functionality of untrustworthy location-based services, a location privacy preserving mechanism (LPPM) can be used to generate a random pseudo-location as a reported location based on a user’s true location. Existing techniques solve an optimization problem on a discrete grid, to construct an optimal LPPM which achieves the highest privacy bounded by minimum tolerable utility, or vice versa. However, they cannot be applied to real-time electric vehicle scenarios since the running time required to generate an optimal LPPM is too long (which can be several days). Another problem deals with optimality of constructed LPPMs. We reveal unexpected cases (anomaly) when the optimal LPPM constructed on a fine grid with superior granularity is worse than that on a coarse one with inferior granularity. We introduce granularity independence as a formal treatment, and propose an optimal LPPM named Divide-and-Coin which can be performed on the fly. Divide-and-Coin improves the running time from at least O (
The Internet of vehicles (IoV) plays an indispensable role in the construction of smart cities, where cars are not just a means of transportation but also a crucial medium for information collection and transmission in the era of big data. With the rapid growth in the volume of data collected from vehicles and the increased awareness of privacy protection, ensuring users’ data security and preventing data breaches in IoV have become an urgent issue to address. Federated learning, as a “data-does-not-move, model-moves” approach, offers a feasible method for protecting user privacy while achieving excellent performance. However, because of the differences of devices, regions and individual habits, data collected from multiple vehicle typically exhibit non-independent and identically distributed (non-IID) characteristics. Traditional federated learning algorithms have slow model convergence when processing non-IID data. In response to this challenge, we propose a multi-stage federated learning algorithm with non-IID data in IoV, named FedWO. In Stage 1, FedWO utilizes the federated averaging algorithm to expedite the global model in reaching a basic level of accuracy. In Stage 2, FedWO employs weighted federated learning, where the weight of a vehicle in the global model is calculated based on its data characteristics. This aggregation results in an improved global model. Moreover, we design a transmission control strategy to reduce communication overhead caused by model transmission. The Stage 3 involves personalized computation, where each vehicle employs its own data for personalized learning, fine-tuning the local model to obtain a model more aligned with local data. We conducted experimental evaluations using a driving behavior dataset. The results demonstrate that, compared with traditional methods, FedWO preserves data privacy while improving the accuracy of algorithms in non-IID data scenarios.
With the development of the Internet of vehicles (IoV), the rapid growth of intelligent vehicles generates a massive amount of data. These data are invaluable for training intelligent IoV application models. Traditional model training requires the centralized collection of raw data through the cloud, consuming substantial communication resources and facing issues like privacy breaches and regulatory constraints. Federated learning (FL) offers a solution by using model transfer instead of data transfer to tackle these challenges. However, practical FL systems are confronted with the issue of malicious users attempting to deceive the server by uploading false local models, known as free-riding attacks. These attacks significantly undermine the fairness and effectiveness of FL. Current research assumes that free-riding attacks are limited to a small number of rational users. However, when there are multiple malicious free-riders, current research falls short in effectively detecting and defending against these attackers. To address this issue, we introduce a novel gradient backtracking based algorithm to identify free-riders. We introduce random testing rounds into standard FL and compare the similarity of user’s gradient between the testing round and the comparison round. It overcomes the challenge of ineffective defense in scenarios involving multiple malicious free-riders. Experimental results on the MNIST and CIFAR-10 datasets demonstrate that the proposed detection algorithm achieves outstanding performance in various free-riding attack scenarios.
The combination of the Internet of vehicles (IoV) and artificial intelligence (AI) has driven the rapid development of autonomous vehicles. Sharing IoV data distributed across different vehicles for training AI models enables more efficient and reliable intelligent driving services. Autonomous vehicles can continuously gather real-time vehicle information, road images and videos among other IoV data, through onboard cameras and sensors. This data are then utilized to optimize and update intelligent traffic models, addressing issues where changes in IoV data result in decreased model accuracy. We propose an efficient and secure federated learning scheme (named as SEFL) for continuous data sharing in an IoV environment to address the problems related to inefficient data collection, catastrophic forgetting problems due to dynamic data updates and privacy leakage from model training parameters. In SEFL, to enable the automatic collection of training samples, each vehicle is based on the global model to only collect IoV data with lower recognition accuracy, and the output with the highest probability is used as the label for that sample. Since vehicle storage space is limited and new samples can overwrite old ones, the data on vehicles are dynamically changing, making traditional fine-tuning training methods prone to catastrophic forgetting. Thus, a dual-knowledge distillation-based training algorithm is proposed in SEFL to ensure that the model learns the knowledge of each sample, maintaining high accuracy. Besides, to prevent privacy leakage from the model parameters between vehicles and servers, an adaptive differential privacy strategy is proposed to achieve client-level privacy protection. Simultaneously, this strategy minimizes the negative impact of differential privacy noise on the accuracy of the global model. Finally, a security analysis and performance evaluation of SEFL scheme are conducted using the GTSRB dataset and vehicle identification dataset. The analysis and experimental results indicate that the proposed SEFL scheme can provide strong privacy protection and efficient data collection. Furthermore, SEFL outperforms existing federated learning-based algorithms in terms of model accuracy.
Vehicular crowdsensing aims to utilize a large number of on-board sensors and computing resources in intelligent vehicles to collect sensing data in a series of areas. Currently, sensing vehicles are usually divided into opportunistic vehicles and participatory vehicles, according to whether the vehicle trajectory can be changed. Here, opportunistic vehicles are with fixed trajectories and cannot be changed. The trajectories of participatory vehicles can be changed based on actual needs. Therefore, selecting the appropriate opportunistic vehicle to complete the sensing task and planning the trajectory of the participatory vehicle is a challenging research problem. We take the different mobility characteristics of the two types of sensing vehicles as the starting point, and manage the two types of vehicles through the crowdsensing platform (CSP). Then, different problems are solved for opportunistic and participatory vehicles, respectively. Specifically, for opportunistic vehicles, it is necessary to select an appropriate vehicle set to complete the sensing task and minimize the cost of CSP. To solve this problem, we propose a reverse auction-based incentive mechanism to select the minimum cost vehicle set to complete the sensing task. It mainly includes two stages: winning vehicle selection and reward payment. Meanwhile, we verify that the proposed method can ensure individual rationality and authenticity. For participatory vehicles, it is necessary to plan the trajectory of each vehicle through CSP scheduling, perform sensing tasks and minimize the cost of CSP. To solve this problem, we propose a deep reinforcement learning-based method to schedule vehicle trajectories and assign different sensing tasks to vehicles. In addition, while minimizing CSP overhead, we also consider the fairness issue of sensing task. The sensing fairness index is introduced to ensure the balance of completion of sensing tasks in different sub-regions. Finally, extensive evaluations based on real-world datasets show that the proposed method performs well and outperforms other benchmark schemes.
Due to the dynamic uncertainty, diversified service types and scarcity of wireless communication resources in the context of vehicle-to-everything, we explore the challenge of ensuring the requirement for multiple quality of service and the effective utilization of wireless resources in the scenario where V2N (vehicle-to-network) and V2V (vehicle-to-vehicle) links coexist and share spectrum in C-V2X (cellular vehicle-to-everything) networks. First, a multi-objective optimization problem is presented to model the decision-making process of channel selection and power control in C-V2X. The problem considers the impact of dynamic changes in the network environment, aiming to make a balance between the performance of the V2V link (i.e., age of information, delay, and capacity) and the capacity of the V2N link. On this basis, V2V wireless resource allocation algorithm based on multi-objective deep reinforcement learning is also proposed for training neural networks to solve the problem. Through the trained neural network model, the Pareto frontier of the multi-objective optimization problem can be obtained. Simulation results demonstrate that the proposed algorithm can achieve the near-optimal performance for different communication links. Compared with four representative algorithms, the age of information for V2V link is reduced by 12.0% to 17.2%, the V2N link capacity is increased by 11.4% to 21.6%, the V2V link transmission success rate is increased by 4.6% to 13.9%, and the decision delay time is reduced by 10.6% to 20.3%.
Efficient and secure data sharing is crucial for the profound application of the intelligent Internet of vehicles, where achieving trusted data sharing between mutually untrusting vehicles has become a major focus of current research. With the characteristics of tamper resistance and traceability, blockchain has emerged as a primary approach to support data circulation in the intelligent Internet of vehicles. Existing blockchain-based data-sharing solutions for the Internet of vehicles suffer from low throughput and security vulnerabilities. In this paper, the blockchain sharding approach is introduced, where a machine learning-based sharding algorithm is utilized to partition road side unit (RSU) with geographical proximity into the same shard and iteratively optimize data-sharing loads within individual shards. This algorithm reduces intra-shard communication latency and subsequently improves throughput while balancing the data-sharing loads among different shards. To prevent bribery attacks within a single shard, a reputation-based intra-shard consensus protocol and the supervisor mechanism are proposed. The proposed protocol involves the election of RSUs with high reputation to participate in the intra-shard consensus process and dynamically calculates the latest reputation of RSUs. High-reputation RSUs are designated supervisors and regularly verify the legitimacy of blocks generated in different shards. Performance evaluation and security analysis demonstrate the scheme enhances the efficiency and security of data sharing in the intelligent Internet of vehicles.
In recent years, the number of software vulnerabilities has increased sharply and its harmfulness has aroused widespread concern in society. Compiling vulnerability utilization code accurately, efficiently and quickly is the key to vulnerability damage assessment and vulnerability repairment. At present, the vulnerability exploitation code mainly relies on manual analysis and writing, which is inefficient. Therefore, how to realize automatic vulnerability exploitation code generation (AEG) is a hotspot and difficulty in software security research field. In this paper, we analyze the representative achievements in this field in recent 30 years. Firstly, we divide the vulnerability automatic utilization process into four typical segments: vulnerability root location, reachable path search, vulnerability primitive generation and utilization code generation. After that we sort out and select the typical work of the above achievements from the three perspectives of human-machine boundary, attack and defense game, and common basic technology. And on this basis, we define the key points, difficulties and phased achievements of the current research. Finally, from the gap between the existing achievements and the practical application of automatic exploit generation, we discuss the bottleneck problems existing in the current research, the future development trend of AEG, and the next research points we should focus on.
Reverse analysis as a key technology plays a vital role in cyber security. It helps analysts gain insight into the behavior of software and vulnerabilities detection, in order to effectively prevent attacks. The growing software scale and complexity urge some research to break down software into modules for rapid analysis via structural and functional information using community discovery algorithms. However, these studies just regard software as a social network consisting of simple nodes and edges missing valuable attribute information. We notice that the contribution of different features to the modular structure of the program is different and varies from samples. Inspired by the innovative application of graph embedding technologies in program analysis, we propose a binary code modularization method called GEBCM. The method transforms an executable program into an attributed graph, and employs graph embedding clustering methods with attention mechanism and ranking mechanism to embed representations and cluster function nodes. The result clusters group binaries into independent parts with more complete functions, revealing the semantic information of complex program structures. Experimental results show that GEBCM outperforms other modularization tools by revealing the original modular layout with an average of 10.2% higher F1 score. Additionally, in the new task of malware decomposition, GEBCM also exhibits better accuracy.
Nowadays, the blockchain ecosystem has been prospering and public blockchain applications and services have been booming. With the concept of Web3.0 getting heated, this trend would be more and more compelling. This fact, while granting more momentum to digital economy development, also brings about a greater challenge to the blockchain supervision practice. The technical traits of blockchain infrastructure and blockchain applications and the need of automatic, multiparty, hierarchical, and multi-dimensional supervision have been calling for innovation and compliance with the supervision schemes designed and implemented for Web2.0. To meet this requirement and face the challenge, we propose a multichain-collaborating, chain-supervising-chain supervision framework. This framework puts forward an architecture consisting of layered and hierarchical supervision chains and heterogeneous supervised chains and designs the basic procedure of blockchain supervision under this architecture. Then we abstract the supervision technologies under this architecture into substrate layer, decision layer, cross-chain layer, access layer, execution layer and data layer. With this abstraction, we propose a series of backbone supervision schemes that are most vital to supervision system construction. Under this framework, we implement a supervision system targeted at blockchain applications and show the effectiveness and feasibility of the system by experiment and exemplary applications.
With the rapid evolution of IPv6 in recent years, the significance of IPv6 network measurement and security analysis has grown substantially. Obtaining a substantial number of active IPv6 addresses has become a fundamental and critical task in this domain. However, the sheer size of the IPv6 address space and the sparsely distributed nature of active hosts present challenges that render brute-force scanning tools, such as ZMap and MASSCAN. While ZMap can scan the entire IPv4 network in just 5 minutes with a 10-gigabit bandwidth, it would take hundreds of millions of years to scan the entire IPv6 network using similar methods. In response to this challenge and in a bid to enhance the efficiency of IPv6-wide scans, researchers have introduced a series of innovative search strategies tailored to IPv6 scans. These strategies aim to enhance the ability to discover assets and mitigate risks within the IPv6 network. We undertake the task of categorizing, organizing, and summarizing the target generation-based scanning approaches proposed by researchers in this field. We conduct a comprehensive analysis, comparing the hit rate, marginal benefit, and time costs of state-of-the-art solutions through real-network scan experiments. Furthermore, we provide valuable insights into the current landscape and emerging trends in IPv6 target generation scanning techniques. By doing so, we contribute to a deeper understanding of IPv6 network analysis and security, ultimately fostering advancements in this critical area of networking research.
As smart terminals and social networks are increasingly integrated into people’s daily life, user behavior identification for social software plays an increasingly important role in network management, network environment supervision, and market research. Social software commonly uses end-to-end encryption protocols for encrypted data transmission, and existing methods usually extract statistical features of the encrypted data for behavior identification. However, these methods have unstable identification performance and require a large amount of data, and these drawbacks affect the practicality of these methods. We propose a social software user behavior identification method for encrypted traffic. First, stable control flow data are identified from the encrypted traffic, and the control service packet payload length sequence is extracted. Then, two neural network models are then designed to automatically extract features from control flow payload length sequences to identify user behavior at a fine granularity. Finally, experiments are conducted with WhatsApp as an example, and the precision, recall, and F1-score of the two neural network models for recognizing WhatsApp user behavior are over 96%. The experimental comparison with similar work proves the stability of the identification performance of the method. In addition, the method can achieve high identification precision with a few control packets, which is of great relevance to the study of real-time behavior identification.
Attackers use the domain names to carry out various kinds of network attacks flexibly. Many scholars have put forward some malicious domain name detection methods based on statistical characteristics and association relationship. However, the two methods have shortcomings in the representation of higher-order relationship of domain name attributes, and cannot accurately present the global higher-order relationship between domains. To solve these problems, a malicious domain name detection method based on embedded feature hypergraph learning is proposed. Firstly, the domain name hypergraph structure is constructed by decision tree based on domain name spatial statistical characteristics. The output of the penultimate node of the decision tree is used as a priori condition to form a hyperedge, and the multi-order correlation between domain name traffic is quickly and clearly represented. Secondly, the character embedding features are enhanced based on the hypergraph structure features, and the hidden higher-order relationships between characters are mined from the domain name data based on the statistical characteristics of domain name space and the encoding characteristics of domain name character embedding. Finally, combined with the real domain name system traffic of China Science and Technology Network, the validity and feasibility are analyzed and evaluated, which can quickly and efficiently detect hidden malicious domain names.
Collaborative learning technologies such as federated learning and swarming learning can fully use distributed data to deeply mine the knowledge contained in the data while protecting data privacy. It has a broad application prospect, especially in the medical and health field, where privacy practices and ethical constraints are emphasized. Collaborative efforts always require reliable participants. The performance of the global model in collaborative learning largely depends on participant selection. However, the existing studies need to pay more attention to the heterogeneity of medical participants’ data. As a result, the performance of the global model, including stability, is difficult to be guaranteed. We propose to solve this problem from the perspective of reputation. Through iterative collaborative learning, reputation participants are selected as much as possible to obtain a stable and reliable high-performance global model in collaborative learning. We first propose an AI medical promise (AMP) to describe a medical institution’s data quality and help form a good AI ecosystem in the medical field. Secondly, an iterative collaborative learning framework based on backward selection (colback-learning) is established. The backward selection method is used to iteratively calculate a well-performing and stable global model in polynomial time complexity to complete AMP calculation and accumulation in a single collaborative learning task. In calculating AMP, a scoring function that comprehensively considers global performance indicators is formulated to guide the training of the global model in the medical field. Finally, using real-world medical data to simulate various collaborative learning scenarios, we have shown that the colback-learning can select reliable participants to obtain a global model with good performance. The model’s performance stability is 1.3 to 6 times higher than that of the state-of-the-art methods. The interpretability of the global model maintains a high consistency with centralized learning.
The metaverse has emerged as a ubiquitous buzzword in the industry, with various sectors eagerly anticipating its potential to unlock forth new opportunities and foster exponential growth, and to serve as a driving force and powerful engine for the digital economy. However, there is a wide divergence of understanding regarding the metaverse across different domains and aspects. This paper aims to provide a framework for a systematic understanding and scientific research of the metaverse through a multidimensional approach. Firstly, we analyze several mainstream metaverse architectural frameworks and their pivotal components, revealing the intricate and multilayered structure of the metaverse. Secondly, we furnish a framework for a meticulous comprehension of metaverse, and propose the metaverse ecosystem architecture, elaborate on the requirements of various levels including the technology system, content system, economic system, collaboration system, and governance system, as well as highlight their interrelationships, and thereby construct the entire metaverse ecosystem. Furthermore, we place special emphasis on the analysis of the crucial role of blockchain technology within the metaverse, particularly underscoring its value in supporting data, digital assets, and collaboration. This study aims to provide a sturdy theoretical foundation and practical guidance for the future development of the metaverse field.
Current serverless computing providers use a coupled resource allocation strategy with low flexibility and a fixed CPU-to-memory allocation ratio. As more types of functions are deployed to the serverless computing platform, the coupled strategy can not satisfy the wide range of resource requirements for these functions. Due to the small granularity of resource allocation and high deployment density in serverless functions, if CPU and memory resource allocation are decoupled, the problem of resource configuration space explosion needs to be solved. In this paper, we present Semi-Share, a decoupled resource manager for serverless functions, which can find the optimal resource configurations for functions while reducing the interference between co-located functions. To solve the resource configuration space explosion problem, Semi-Share builds a two-layer resource allocation architecture, which divides the resource configuration space into multiple subspaces to reduce problem complexity. The first layer is the function cluster, which is based on the resource preference and historical load information of the functions. The resource configuration space is divided according to these clusters. The second layer is resource allocation, which leverages the Bayesian optimization and weighted scoring function to guide Semi-Share to search in the right direction in the configuration space and reduce the time overhead. The experimental results show that Semi-Share greatly reduces the search time of searching the optimal resource configuration by using the two-layer architecture, reduces the average configurations sample by 85.77% compared with the widely used gradient descent search method, and improves the function performance by 42.72% on average. Compared with COSE, a coupled resource allocation system that also uses Bayesian optimization, Semi-Share can improve the function performance by 32.25% on average.