Salient object detection plays a pivotal role in the interpretation of remote sensing imagery. However, this task remains challenging due to the inherent complexities of remote sensing scenes, which include cluttered backgrounds, substantial scale variations, indistinct object boundaries, and environmental noise. To improve localization accuracy in these complex scenarios, we propose a novel edge-guided stratified fusion network (ESFNet) for accurate remote sensing salient object extraction. The network achieves stratified fusion of fine-grained edge cues and rich semantic information through its hierarchical architecture, enabling precise salient object localization. Specifically, we design an EMFM to explicitly incorporate contour semantics, thereby enhancing boundary delineation while effectively suppressing background clutter and noise interference. Moreover, a stratified attention module aggregates contextual information across multiple semantic levels using a hierarchical attention mechanism. This dual-path architecture facilitates synergistic interaction between edge precision and semantic coherence, significantly improving feature discriminability and spatial consistency. Extensive experiments on two benchmark datasets demonstrate that our ESFNet achieves state-of-the-art performance, exhibiting remarkable superiority in detecting salient objects with sharp boundaries and consistent semantics across diverse remote sensing scenarios.

In the current context of the rapid development of artificial intelligence, the security protection of deep learning models has become an important issue that needs to be solved urgently. Among them, adversarial sample attacks, as a significant threat, can mislead model decisions through small but well-designed perturbations, seriously damaging the reliability and robustness of the system. To this end, this paper proposes a novel defense strategy called adaptive masked autoencoder defense (AMAEDefense), which is committed to building a more stable machine learning ecosystem. The core idea of AMAEDefense is to combine adaptive masking technology and autoencoder architecture to implement refined filtering and reconstruction operations on input data to eliminate potential malicious interference signals. First, the adaptive masking mechanism is used to locate and shield critical areas containing attack components. Then, the hidden content is recovered with the help of a highly optimized autoencoder network, and finally, the purified clean version is output for downstream tasks. In order to verify the effectiveness of the proposed method, an exhaustive experimental evaluation was carried out on multiple benchmark datasets. The results show that AMAEDefense shows excellent defensive performance in the face of various mainstream attacks such as FGSM, PGD and CW. On the CIFAR-10 dataset, the successful defense rate of AMAEDefense is as high as 94%. To ensure the defense effect, the quality loss of image reconstruction is controlled at a low level and the quality retention rate is above 86%. Compared with other competitive schemes, the running time of this research method is shortened by about 30%, reflecting the advantage of higher computational efficiency. AMAEDefense significantly enhances the ability of deep learning models to resist adversarial samples and considers the balance between execution speed and output accuracy, providing a valuable reference path for subsequent related research.

Transfer-based black-box attacks are an important tool for evaluating deployed vision models, yet adversarial examples generated from Vision Transformer (ViT) surrogates often exhibit limited cross-architecture transferability. Existing momentum-based attacks are effective for convolutional neural network (CNN) surrogates, but they can accumulate stale directions and overfit the surrogate when the source model is a ViT. This paper presents Ada-MGNS, a ViToriented transferable attack that combines adaptive momentum with deep attention guidance. The adaptive component measures the directional discrepancy between the current guided gradient and the accumulated trajectory, and then attenuates stale momentum when the search direction becomes unstable. The guidance component fuses the classification gradient with an auxiliary gradient extracted from the last transformer block’s attention responses, encouraging perturbations to disturb both output decisions and semantic aggregation. Experiments on ImageNet with four ViT surrogates, thirteen standard black-box targets, and five defense models show that Ada-MGNS consistently improves attack success rates over representative ViT-specific baselines, remains compatible with DI/TI transformations and effective against adversarially trained and purification-based defenses.

Due to the challenges of synthetic aperture radar (SAR) data acquisition and the high cost of manual annotation, utilizing labeled optical images to learn from unlabeled SAR data has received great attention. Cross-modal domain adaptation (DA) from optical to SAR imagery presents a particularly difficult problem because of the inherent modality gap between these two imaging paradigms. To address the problem of unsupervised ship detection in SAR images, we conduct domain adaptation experiments from ship images in the DIOR dataset to the SSDD dataset. However, traditional domain adaptation methods are insufficient to address the significant modality differences between optical and SAR images. Although the unconstrained feature alignment strategy is effective between domains with small differences, it inadvertently expels SAR features from the supervised recognition space, ultimately reducing detection performance. To mitigate this issue, we propose a new framework, DAIR, which integrates an innovative inversion regularization module (IRM) and a task-correlation enhancement (TCE) strategy to improve domain adaptation. Specifically, IRM acts as a feature-space regularizer to counteract deviation caused by aggressive alignment, while TCE explicitly models task interdependency to alleviate the effects of task independence. Evaluated on the DIOR and SSDD datasets, our method achieved improvements of 5% and 8% in AP50 and APm, respectively, over the baseline method DA Faster R-CNN. In few-shot scenarios, our method attained gains of 14.1%, 10.6%, and 15.4% in AP50 compared to the state-of-the-art method under 3-shot, 5-shot, and 10-shot settings, which demonstrates stronger generalization ability with limited data. Source code and models are available at https://github.com/whatbb/DAIR/tree/main

The prevalence of malware in recent years has led to many major information security incidents. Tens of millions of computers around the world are infected by malware and cause significant losses to individuals and businesses. Due to the diversity of malware, it is difficult to extract features from malware, which make malware detection not conductive to the application of AI technology. Malware detection can get help from graph neural network (GNN) to learn the characteristics of malware. Zero-Shot learning (ZSL) can transfer knowledge from known classes to unknown ones and facilitate malware recognition without large samples effectively. The structure and semantic embedding of malware graphs in the knowledge graph, as well as the use of correlations between malware categories or semantic vectors of categories, improve the classification accuracy of ZSL and GNN models for unknown malware and variant malware. In this paper, the graph convolutional network (GCN), graph attention network (GAT) and ZSL are employed for malware detection. To evaluate the effectiveness of GCN and GAT model, the modes are evaluated with accuracy, precision, and recall. Experimental results show that the GCN and GAT models can detect malware effectively.

Introduction Due to the complex environment of the substation, the inspection work of the substation becomes time-consuming and laborious. As a result, the substation inspection robot has gradually become a hot research point. Objectives At present, the mainstream intelligent inspection robots for substations use high-precision LiDAR sensors for navigation, which has high navigation accuracy but cannot identify the types of obstacles and the road contour boundary, seriously affecting the inspection performance and efficiency. Meanwhile, the high-precision 3D laser radar is too expensive to afford. In order to solve these problems, a novel navigation assistant method is proposed in this paper, which is based on multisensory information fusion. Methods Asynchronous information matching with multiple sensors was used to match the information collected by different sensors to deal with the time asynchrony. To express the height of obstacles, 2D laser radar was applied to create 3D imaging by being combined with inertial measurement unit (IMU). For perceiving and under-standing the substation environment independently by the inspection robot, ENet was given over to segment color point cloud maps, which was built by introducing optical sensor data. The method was implemented based on the ROS system and transplanted to embedded platform of the inspection robot. Result Finally, experimental results show that, compared with VLP-32C 3D laser radar sensors, the data volume of navigation assistance module had reduced by 95%. Meanwhile, after training the ENet network, the mean average accuracy value had achieved 86%, which meets the needs of practical engineering applications. Conclusion In addition, the inspection robot equipped with the navigation assistance module is successfully tested in multiple substations, which shows that the robot can not only identify the road contour and the type of obstacles on the way, but also reduce the amount of data and the cost of hardware.

The rapid proliferation of ransomware variants necessitates more effective detection mechanisms, as traditional signature-based methods are increasingly inadequate. These conventional methods rely on manual feature extraction and matching, which are time-consuming and limited to known threats. This study addresses the escalating challenge of ransomware threats in cybersecurity by proposing a novel deep learning model, LSTM-EDadver, which leverages Generative Adversarial Networks (GANs) and Carlini and Wagner (CW) attacks to enhance malware detection capabilities. LSTM-EDadver innovatively generates adversarial examples (AEs) using sequential features derived from ransomware behaviors, thus training deep learning models to improve their robustness and accuracy. The methodology combines Cuckoo sandbox analysis with conceptual lattice ontology to capture a wide range of ransomware families and their variants. This approach not only addresses the shortcomings of existing models but also simulates real-world adversarial conditions during the validation phase by subjecting the models to CW attacks. The experimental results demonstrate that LSTM-EDadver achieves a classification accuracy of 96.59%. This performance was achieved using a dataset of 1328 ransomware samples (across 32 ransomware families) and 519 normal instances, outperforming traditional RNN, LSTM, and GCU models, which recorded accuracies of 90.01%, 93.95%, and 94.53%, respectively. The proposed model also shows significant improvements in F1-score, ranging from 2.49% to 6.64% compared to existing models without adversarial training. This advancement underscores the effectiveness of integrating GAN-generated attack command sequences into model training.

Purpose The purpose of this paper is to use the corresponding magnetic sensor and detection method to detect and image the defects of small diameter pipelines. Urban gas pipeline is an energy transportation tool for urban industrial production and social life, which is closely related to urban safety. Preventing the occurrence of urban gas pipeline transportation accidents and carrying out pipeline defect detection are of great significance for the urban economic and social stability. To perform pipeline defect detection, the magnetic flux leakage internal detection method is generally used in the detection of large-diameter long-distance oil and gas pipelines. However, in terms of the internal detection of small-diameter pipelines, due to the heavy weight, large structure of the detection device and small pipe diameter, the detection is more difficult. Design/methodology/approach In order to solve the above matters, self-made three-dimensional magnetic sensor and three-dimensional magnetic flux leakage imaging direct method are proposed for studying the defect identification. Firstly, for adapting to the diameter range of small-diameter pipelines, and containing the complete information of the defect, a self-made three-dimensional magnetic sensor is made in this paper to improve the accuracy of magnetic flux leakage detection. And on the basis of it, a small diameter pipeline defect detection system is built. Secondly, as detection signal may be affected by background magnetic field interference and the jitter interference, the complete ensemble empirical mode decomposition with adaptive noise method is utilized to screen the detected signal. As a result, the useful signal is reconstructed and the interference signal is removed. Finally, the defect contour inversion imaging of detection is realized based on the direct method of three-dimensional magnetic flux leakage imaging, which includes three-dimensional magnetic flux leakage detection data and data segmentation recognition. Findings The three-dimensional magnetic flux leakage imaging experimental results shown that, compared to the actual defects, the typical defects, irregular defects and crack groove defects can be analyzed by the magnetic flux leakage defect contour imaging method in qualitative and quantitative way respectively, which provides a new idea for the research of defect recognition. Originality/value A three-dimensional magnetic sensor is made to adapt the diameter range of small diameter pipeline, and based on it, a small-diameter pipeline defect detection system is built to collect and display the magnetic flux leakage signal.

Identifying wildfires is a key task to ensure timely and effective response and prevent the spread of wildfires. The widely used methods for identifying wildfires based on infrared and satellite remote sensing still have problems such as poor identification accuracy and high deployment costs. To address the high deployment costs, this paper adopted the method of deep learning network identifying visible light images.To address the issue of poor accuracy, this paper proposed an intelligent wildfire identification method based on Weighted Boxes Fusion (WBF) and Convolutional Block Attention Module (CBAM). In order to enhance the feature expression of flames and smoke, this paper adopted two different scale SSD networks combined with CBAM to adjust the feature weights of relevant regions and channels adaptively and uses them as the fusion base model. In order to further improve the performance of the model and integrate two different feature extraction strategies, this paper adopted the WBF method to weigh and fuse the identification results of the base model. The method was evaluated on a self-built dataset. The experimental results showed that the mAP of this method reached 96.19%, which is the optimal performance method compared to existing methods. The model proposed in this paper can effectively detect wildfires from images and has practical application potential in wildfire monitoring and management.

The emergence of various intelligent mobile applications necessitates the deployment of powerful deep learning models on resource-constrained devices. Device-edge co-inference offers a promising solution by allocating neural networks. It is necessary to balance the computation and communication cost by compressing intermediate features. Current methods for feature compression usually separate the inference task from communication design and have not yet considered the actual impact of wireless channels on feature compression. In this paper, we propose an end-to-end channel-adaptive feature compression approach to achieve efficient feature compression under wireless channels. Additionally, in order to fulfill human perception requirements, we propose a mirror model based on feature compression, aiming to restore images with the original resolution from compressed features. We conduct comprehensive experiments to validate the effectiveness of the proposed method.

A novel acquisition scheme of pseudo-noise (PN) codes is proposed for spreading satellite communication systems relying on the proposed multi-voter model of this paper. Based on the proposed model, the acquisition of PN codes can be attributed to a voting and selecting mechanism to pick out the erroneous chips. Although message passing algorithm (MPA) is a feasible algorithm for decoding PN codes, MPA does not get good detection performance due to the limited number of parity check equations in the acquisition scheme. To overcome the negative impact of the limited number of parity check equations, this paper proposes single-voter and multi-voter models, which combine chip and sequence estimation on the basis of the chip-flipping (CF). It is known that the bit-flipping (BF), weighted-bit-flipping (WBF) and other algorithms based on BF are credible for low density parity check (LDPC) codes. Because of the lack of research, these algorithms have not been extended to the detection of PN codes. As the same as BF, the inputs of the proposed CF algorithm are the hard-decision samples. For the proposed CF, there exists an optimal flipping-threshold which is similar to the random weight of the WBF. Owing to the low computation complexity of CF, the unlimited number of parity check equations can be enabled in the voting model. The experimental results show that the detection performance of the proposed method of $N > $ 15 is improved by 2 dB compared with MAP-based method and 4 dB compared with LEAP-based method at the detection probability 99%.

Download This Paper Open PDF in Browser Add Paper to My Library Share: Permalink Using these links will ensure access to this page indefinitely Copy URL Copy DOI

A novel acquisition scheme of pseudo-noise (PN) codes is proposed for spreading satellite communication systems relying on the proposed multi-voter model of this paper. Based on the proposed model, the acquisition of PN codes can be attributed to a voting and selecting mechanism to pick out the erroneous chips. Although message passing algorithm (MPA) is a feasible algorithm for decoding PN codes, MPA does not get good detection performance due to the limited number of parity check equations in the acquisition scheme. To overcome the negative impact of the limited number of parity check equations, this paper proposes single-voter and multi-voter models, which combine chip and sequence estimation on the basis of the chip-flipping (CF). The experimental results show that the detection performance of the proposed method of N =30 is improved by 2 dB compared with MAP-based method and 4 dB compared with LEAP-based method at the detection probability 99%.

The prevalence of malicious software (Malware) in recent years has led to many major information security incidents. Tens of millions of computers around the world are infected by malware and cause significant losses to individuals and businesses. RaaS (Ransomware-as-a-Service) is a new type of cybercrime model. Ransomware has caused losses exceeding billions of dollars every year, making it a major threat to network security. Due to the diversity of ransomware, it is difficult to extract features from ransomware, which make ransomware detection not conductive to the application of AI technology. Ransomware detection can get help from graph neural network (GNN) to learn the characteristics of ransomware. In this research, the graph convolutional network (GCN) and graph attention network (GAT) are employed for malware detection, and compare their performance with each other. Cuckoo Sandbox is deployed to log malicious behaviors generated by ransomware, and the JSON report generated by Cuckoo Sandbox is employed to extract API call sequences for ransomware detection. To evaluate the effectiveness of GCN-based and GAT-based model, the modes are evaluated with ransomware samples downloaded from Malware Bazaar Database and examined with accuracy, precision, recall and ROC curve. Experimental results show that the GCN-based and GAT-based models can detect ransomware effectively. The GCN-based model reaches better results than the GAT-based model.

Recently, pixel-value-ordering (PVO) based reversible data hiding (RDH) methods have become hotspots in spatial domain RDH research. In PVO-based methods, the pixel value correlation within a local region is exploited for data embedding. To embed secret data, the cover images are partitioned into non-overlapped rectangle blocks to be sorted locally, and a lower embedding distortion is thus obtained. However, their embedding capacities are often limited by the fixed local blocks from which pixels are sorted. To maintain the advantages and overcome disadvantages of the PVO-based methods, a new global pixel-value-ordering (GPVO) framework is proposed in this paper. By applying our GPVO framework, a PVO-based method can utilize pixels from any position in the cover image rather than from local pixel blocks. Then, through the dynamic sequence partition realized by the GPVO framework for the first time, the embedding capacity of PVO-based methods can be greatly improved while reducing embedding distortion further reduced. Finally, we propose a two-stage pairwise embedding scheme to be applied to the sequences, aiming to achieve an advanced embedding performance. Experimental results illustrate that the proposed GPVO works better than other related state-of-the-art RDH methods.

With the development of e-commerce, e-commerce security incidents have caused information security to attract the attention of the general public. According to statistics, a large number of computers are infected by malware such as Trojans and zombies, and are used by the botnet. Malware can invade computer systems and exploit system or network vulnerabilities for intrusion. For malware detection, machine learning algorithms are used to detect malware, but they cannot incorporate background knowledge. Malware detection needs to be integrated with knowledge graphs, deep learning, and reinforcement learning. In this study, the proposed model was developed by employing deep learning, graph neural network, knowledge graph, and reinforcement learning to detect malware. To evaluate the effectiveness of the proposed model, the proposed model was evaluated with malware samples collected from MalwareBazaar, DikeDataset, and VirusShare and examined with accuracy, precision, and recall. Experimental results showed that the proposed model detected malware effectively.

Face recognition technology has been widely used in the field of artificial intelligence. The technology needs to be carried out normally under the appropriate light, however, there is not ideal light, even poor-lighted for the face recognition device, and with the head in deflect angle. The poor-lighted under various head poses will influence the face recognition significantly. To address the issue, we present a novel and practical architecture based on deep fully convolutional neural network and generative adversarial networks for illumination normalization of facial images. The proposed method is termed as illumination normalization generative adversarial network. Compared with previous methods based on deep learning, our approach does not require identity and illumination label as input. We preserve identities of faces by an elaborately-designed generator together with content loss. Moreover, the framework of our scheme is simpler than previous methods based on deep learning. It can address the illumination of frontal and non-frontal face. In order to fairly evaluate the proposed method against state-of-the-art models, the peak signal to noise ratio is employed to estimate the performance of illumination normalization algorithm. Experimental results show that the proposed method achieves favorable normalization results against previous models under various head poses and illumination challenges.

Pan-sharpening can provide multispectral images with high spatial resolutions, which are useful for many remote sensing image applications. Currently, deep learning technology has been widely used in pan-sharpening. Most of these deep learning-based methods ignore domain-specific knowledge that can improve spatial and spectral performance. Some improved methods adopt an injection structure which injects initial details obtained from panchromatic images into multispectral images through detail mapping. However, the initial details lack frequency-domain information. Moreover, the detail mapping is completed by convolutional neural networks, which lack sufficient nonlinearity to generate rich and diverse details. To solve the above problems, a domain-specific knowledge-driven pan-sharpening framework based on a detail injection structure is proposed, which includes two stages of knowledge-driven initial detail acquisition and data-driven detail mapping. In the first stage, in order to perform better feature reconstruction in the frequency domain, the PAN-MS method is introduced to provide initial details containing frequency-domain information. In the second stage, a newly designed detail-mapping generative adversarial network (GAN) maps initial details to more various output details. Experiments conducted on three public datasets has proven that the proposed algorithm outperforms some state-of-the-art methods in terms of spatial and spectral performance.

This paper presents an effective method for strengthening the discriminative ability of high-level deep features by enhancing and aggregating discriminative part-level features for the fine-grained vehicle recognition task. In general, the task of visual recognition concentrates more on the visual differences at the object level. However, for fine-grained object recognition, the visual differences between target objects typically exist in local discriminative areas, so it is more concerned about extracting fine-grained features from these part regions. In this context, we propose solving this issue with a novel feature extraction method from two perspectives: the generation of more feature descriptors of part regions through the learning process of deep networks and the aggregation of part-level discriminative features. This approach is designed to improve the backbone networks to generate finer-level part features through a part-level feature enhancement module and to investigate the intrinsic part-level features of the backbone networks with the help of a feature aggregation module. The enhancement module efficiently finds the finer features highly correlated to the part regions. Then the feature aggregation module builds correlations of similar part features through feature grouping and fusion. Moreover, our proposed method does not require additional parts annotations and achieves comparable performance on two widely-used benchmarks for recognizing fine-grained vehicle types. Experimental results and explainable visualizations demonstrate the effectiveness of the proposed method.

Recently, there has been a surge of interest and attention in Transformer-based structures, such as Vision Transformer (ViT) and Vision Multilayer Perceptron (VMLP). Compared with the previous convolution-based structures, the Transformer-based structure under investigation showcases a comparable or superior performance under its distinctive attention-based input token mixer strategy. Introducing adversarial examples as a robustness consideration has had a profound and detrimental impact on the performance of well-established convolution-based structures. This inherent vulnerability to adversarial attacks has also been demonstrated in Transformer-based structures. In this paper, our emphasis lies on investigating the intrinsic robustness of the structure rather than introducing novel defense measures against adversarial attacks. To address the susceptibility to robustness issues, we employ a rational structure design approach to mitigate such vulnerabilities. Specifically, we enhance the adversarial robustness of the structure by increasing the proportion of high-frequency structural robust biases. As a result, we introduce a novel structure called Robust Bias Transformer-based Structure (RBFormer) that shows robust superiority compared to several existing baseline structures. Through a series of extensive experiments, RBFormer outperforms the original structures by a significant margin, achieving an impressive improvement of +16.12% and +5.04% across different evaluation criteria on CIFAR-10 and ImageNet-1k, respectively.