As an extension of cloud computing, fog computing (FG) deploys storage resources to the edge of the network to reduce latency and improve real-time processing capabilities. However, the distributed architecture and complex network environment of FG make it vulnerable to abnormal behaviors. Therefore, it is necessary to design practical anomaly detection methods to protect the security of fog nodes. Previous research mainly focuses on static anomaly detection methods, which use offline training to protect fog nodes. However, these static methods make it difficult to cope with concept drift scenarios in fog environments where both threat situations and normal behaviors are constantly changing. Furthermore, existing dynamic methods do not consider the storage space of fog servers and the labeling limitations of experts. To address the above challenges, we propose an online adaptive anomaly detection framework (OAADF), which consists of three key modules: a statistics-based dynamic feature selection module for removing redundant features, a drift detection module based on block similarity changes for identifying concept drift, and a score-based update module for model adaptation. Experimental validation using the CICIDS2017 and 5G-NIDD datasets in the NS-3 simulation environment demonstrates the superior performance and practicality of OAADF, surpassing the state-of-the-art (SOTA) solutions.

Abstract Anomaly detection in the Controller Area Network (CAN) bus is critical for ensuring the security and reliability of intelligent connected vehicles, which are increasingly prevalent. While existing anomaly detection strategies offer some benefits, they often face challenges such as limited feature extraction and data imbalance, which reduce their effectiveness. To address these issues, in this paper, we propose an unsupervised intrusion detection method based on CAN message graph named CGTS. Specifically, we first construct a message graph based on CAN message sequences. A Graph Transformer is then employed to extract complex structural information, accurately capturing the intrinsic connections between messages. Furthermore, to address the data imbalance problem, we integrate the Support Vector Data Description algorithm after the Graph Transformer model. This algorithm identifies anomalous behaviors efficiently without relying on a priori labels. Experiments conducted on public datasets, including Car-Hacking and CAN-Train-and-Test, demonstrate the efficacy of CGTS. The model achieves an average accuracy exceeding 0.990, precision above 0.995, and an F1-score nearing 0.993. These results highlight CGTS can effectively detect multiple injection attacks and significantly improve the CAN bus intrusion detection performance.

In recent years, cyberattacks against automobiles have exposed significant security threats to in-vehicle networks. The vulnerability of communication signals to malicious interference and manipulation can lead to serious system failures or abnormal behavior. The existing in-vehicle network intrusion detection methods do not fully exploit the time-frequency characteristics of controller area network (CAN) signals. This limitation reduces their effectiveness in capturing subtle changes and signal complexity. Based on the above motivation, from the perspective of signal perception, we propose an entropy-driven, time-frequency analysis framework for in-vehicle network intrusion detection. The framework integrates a signal sampler, a frequency-domain detector, and a time-domain detector. The signal sampler, as the system’s front-end module, extracts real-time physical signal data streams from CAN messages. The frequency-domain detector identifies frequency components, detecting high-frequency disturbances and cyclic variations. Meanwhile, the time-domain detector captures instantaneous changes and sudden anomalies. It analyzes signal complexity and anomalies through both stream and block detection. Experimental results demonstrate that the proposed method performs well under various attack scenarios, offering superior detection and real-time performance. It effectively senses multiple signal anomalies, providing a robust intrusion detection solution for modern in-vehicle networks.

The remarkable progress in cross-modal retrieval relies on accurately annotated multimedia datasets. In practice, most existing datasets used for training cross-modal retrieval models are automatically collected from the Internet to reduce data collection costs. However, it inevitably contains mismatched pairs, i.e., noisy correspondences, thus degrading the model performance. Recent advances utilize the predicted similarity distribution of individual samples for noise validation and correction, which easily faces two challenging dilemmas: (1) confirmation bias and (2) unstable performance with increasing noise. In light of the above, we propose a generalized Bias Mitigation and Representation Optimization (BMRO) framework. Specifically, we propose a Bias Estimator (BE) to estimate the unbiased confidence factor of a sample by contrasting it against its nearest neighbors. Unbiased confidence factor can precisely adjust sample contribution and enhance accurate sample division. This facilitates the Adaptive Representation Optimizer (ARO) in providing tailored optimization strategies for clean and noisy samples. ARO performs contrastive learning between clean samples and generated hard samples, thus promoting the generalizability and robustness of the representation. Besides, it utilizes complementary learning to reduce incorrect guidance from noisy samples. Extensive experiments on five visual-text benchmarks verify that our BMRO can significantly improve the matching accuracy and performance stability against noisy correspondences.

Text-based Person Retrieval (TPR) aims to retrieve relevant images of specific pedestrians based on the given textual query. The mainstream approaches primarily leverage pretrained deep neural networks to learn the mapping of visual and textual modalities into a common latent space for cross-modality matching. Despite their remarkable achievements, existing efforts mainly focus on learning the statistical cross-modality correlation found in training data, other than the intrinsic causal correlation. As a result, they often struggle to retrieve accurately in the face of environmental changes such as illumination, pose, and occlusion, or when encountering images with similar attributes. In this regard, we pioneer the observation of TPR from a causal view. Specifically, we assume that each image is composed of a mixture of causal factors (which are semantically consistent with text descriptions) and non-causal factors (retrieval-irrelevant, e.g., background), and only the former can lead to reliable retrieval judgments. Our goal is to extract text-critical robust visual representation (i.e., causal factors) and establish domain invariant cross-modality correlations for accurate and reliable retrieval. However, causal/non-causal factors are unobserved, so we emphasize that ideal causal factors that can simulate causal scenes should satisfy two basic principles:1） Independence: being independent of non-causal factors, and 2）Sufficiency: being causally sufficient for TPR across different environments. Building on that, we propose an Invariant Representation Learning method for TPR (IRLT), that enforces the visual representations to satisfy the two aforementioned critical properties. Extensive experiments on three datasets clearly demonstrate the advantages of IRLT over leading baselines in terms of accuracy and generalization.

Text input is one of the common interaction tasks in virtual environments. However, current inputting methods (e.g., laser-input: aim-and-shoot technique) have many limitations, such as inefficiency, lack of precision, and fatigue of long-text inputting. We propose a Climbing Keyboard method to allow easier, faster, and more accurate text input, and use tilt instead of precise aiming. The selected target changes from a specific letter to a group of letters with such a tilt-based interaction based on the QWERTY layout, which aims to reduce the learning cost, especially for novice users. Meanwhile, expert users can focus on the screen without looking at the keyboard. We designed three user studies to evaluate the performance of proposed method, including the verification of the usability of the tilt interaction method in the first study, optimization of the tilt angle range in the second study, and evaluation of the learning curve of Climbing keyboard in the last study. Our results showed that participants can reach 16.48 words per minute after an hours of training.

Abstract Recently, existing FCNs‐based methods have shown their advantages in processing object boundaries. However, these methods still suffer from false object interference, which appears in saliency predictions. To solve this problem, an edge‐interior feature fusion (EIFF) framework is proposed, which consists of an internal‐boundary decoupled generation structure with receptive field enlargement and attention mechanism enhancement, and a salient feature refinement module. Specifically, the framework first learns edge features and interior features through an internal‐boundary decoupling generation network, which is supervised by labels obtained by decoupling ground‐truth through an image erosion algorithm. Then, feature refinement module (FRM) is designed to purify the coarse prediction by focusing on the ambiguous regions through a mining strategy to generate the final saliency map. To compensate for shortcomings of the BCE and IU loss, we also introduce a weighted loss to guide our model to focus more on the error‐prone parts. Experimental results on five benchmark datasets demonstrate that the proposed method performs favorably against 19 state‐of‐the‐art approaches under four standard metrics.

Seismic data denoising has gained much attention from scholars as a crucial part of seismic data processing. With the development of deep learning technology, numerous algorithms well employed in natural image denoising have been used for seismic data denoising. However, seismic data are crucially an array of geophone vibration signals, which have numerous unique structural properties compared with natural images. Thus, employing the traditional image algorithms rather than those developed specifically for seismic data is inadequate to extract all seismic features. Additionally, compared with natural images, it is difficult to acquire noise-free ground truth, which restricts the use of supervised deep learning approaches in seismic data denoising. To this end, a self-supervised inter-trace seismic data denoising network (STSNet) that requires only one seismic section for random noise attenuation is proposed. Furthermore, we adopt a single trace vibration signal as the basic unit and fully consider the characteristics of seismic signals. This is the first time to introduce a self-learning spatial attention mechanism among seismic traces to focus on the noise components, which prompts the network’s fitting performance. Different comparative experiments demonstrate that our approach can achieve exceptional denoising performance even if only observing a single seismic section. Additionally, the ablation experiments also confirm the efficiency of spatial attention.

Neural network-based approaches have taken the lead in medical image segmentation with the encoder-decoder architecture. However, these approaches are still limited to one neural structure, which is short in leveraging the strengths of the three dominant structures (Convolutional Neural Network, Transformer, and Multilayer Perceptron) simultaneously. Furthermore, simple skip connections cannot effectively bridge the semantic gap between the encoder and decoder at the same level. To alleviate the above problems, this paper proposes Mixed-Net,haode which cleverly formulates a strategy to synergize three neural network structures for medical image segmentation. Specifically, our method innovatively designs two components, namely a Semantic Gap Bridging Module (SGBM) and a Global Information Compensation Decoder (GICD). Convolution-based SGBM can validly expand the receptive field and combine shallow and high-level representations by replacing original skip connections. Equally importantly, we present a GICD containing convolution and transformer, which can adequately incorporate local refinement features and global representations in the information decoding space. We evaluate Mixed-Net on 3 different medical image segmentation datasets. Surprisingly, our method sets the new state-of-the-art performance and demonstrates stronger generalization capability.

The introduction of proportional-integral-derivative (PID) controllers into cooperative collision avoidance systems (CCASs) has been hindered by difficulties in their optimization and by a lack of study of their effects on vehicle driving stability, comfort, and fuel economy. In this paper, we propose a method to optimize PID controllers using an improved particle swarm optimization (PSO) algorithm, and to better manipulate cooperative collision avoidance with other vehicles. First, we use PRESCAN and MATLAB/Simulink to conduct a united simulation, which constructs a CCAS composed of a PID controller, maneuver strategy judging modules, and a path planning module. Then we apply the improved PSO algorithm to optimize the PID controller based on the dynamic vehicle data obtained. Finally, we perform a simulation test of performance before and after the optimization of the PID controller, in which vehicles equipped with a CCAS undertake deceleration driving and steering under the two states of low speed (≤50 km/h) and high speed (≥100 km/h) cruising. The results show that the PID controller optimized using the proposed method can achieve not only the basic functions of a CCAS, but also improvements in vehicle dynamic stability, riding comfort, and fuel economy.

As FlexRay communication protocol is extensively used in distributed real-time applications on vehicles, signal scheduling in FlexRay network becomes a critical issue to ensure the safe and efficient operation of time-critical applications. In this study, we propose a rectangle bin packing optimization approach to schedule communication signals with timing constraints into the FlexRay static segment at minimum bandwidth cost. The proposed approach, which is based on integer linear programming (ILP), supports both the slot assignment mechanisms provided by the latest version of the FlexRay specification, namely, the single sender slot multiplexing, and multiple sender slot multiplexing mechanisms. Extensive experiments on a synthetic and an automotive X-by-wire system case study demonstrate that the proposed approach has a well optimized performance.

Gear-position-decision (GPD) tactics strongly affect the performances of automatic transmissions (AT) and, therefore, the performance of the vehicle. Since the electronic control methods were introduced into ATs, many advanced techniques have been raised to make AT vehicles more human friendly and better in fuel economy and dynamic behaviors. As a type of emerging AT, the automated manual transmissions (AMT) are being researched and developed in all relevant technologies. In this paper, we proposed a driving knowledge-based GPD (KGPD) method for AMTs. The KGPD algorithm is composed of a driving environments and driver's intentions estimator, the shift schedules for each typical driving environment and driver's intention situations, and an inference logic to determine the most proper gear position for the present situation. The estimator identifies the driving environments and features of driver's intentions, which are divided into some typical patterns. Based on the identified results, the gear-position inference algorithm calculates the best gear position at the moment. In fact, the method just simulates the course of a driver's making gear-position decision when driving an automobile with manual transmission. The test results show that the AMT with the method gives less unnecessary shifting, conducts more proper gear positions, and behaves better in subjective assessment than that with the method that is directly based only on automotive state parameters.

A gear position decision method used in automated mechanical transmission (AMT) is introduced. The algorithm of the method is composed of a driving environments and driver's intentions estimator, the shift schedules suit each typical driving environment and driver's intention situation, and an inference logic to determine the most proper gear position for the present situation. The estimator identifies the driving environment and features of driver's intentions, which are divided into some typical patterns. Based on the identified results, the gear position inference algorithm calculates out the best gear position at the moment. The method just simulates the course of a driver making a gear position decision when driving an automobile with manual transmission. The test results show that the automated mechanical transmission with the method gives less unnecessary shifting and more proper gear positions than that with the shift schedule algorithms, which calculate the gear positions only based on the automotive state parameters.