In dynamic environments, large language models need to keep adapting to new tasks, but continual learning often suffers from forgetting, limited transfer, and vulnerability to adversarial perturbations. To address this, we present AdvCL, which repurposes adversarial perturbations as a geometric control signal for stable continual adaptation. AdvCL combines three plug-in modules: Intra-Smooth promotes local smoothness via small adversarial perturbations; Proto-Clip uses similarity clipping to prevent excessive alignment to current task prototype; and Inter-Align applies directional alignment toward previous task prototype to reduce representational gaps. Experiments show consistent gains in both standard performance and robustness, with lower forgetting and stronger transfer. We further analyze key mechanisms by quantifying the sensitivity of Intra-Smooth to perturbation settings and the effect of Inter-Align on task similarity and geometric distance. In summary, the modules provide complementary gains when combined, and each can also be integrated individually into diverse CL paradigms, including replay, regularization, and dynamic architectures, thereby offering a geometric control mechanism for continual learning.

With the widespread popularity of social media, sarcastic expressions have become increasingly common in online communication. Accurately detecting such expressions is crucial for understanding user intentions, conducting sentiment analysis, and maintaining a healthy online environment. However, sarcasm detection faces numerous challenges, such as the subtlety of text semantics and the complexity of multi-modal information fusion. To effectively address these issues, this paper proposes a sarcasm detection method that integrates Large Language Models (LLMs) with a cross-modal adaptive attention network. In terms of text processing, the powerful language capabilities of large language models are utilized to perform in-depth semantic representation of the input text, capturing rich language patterns, contextual relationships, and implied semantic information, thereby providing high-quality textual features. Meanwhile, considering the close association between multi-modal information and text, a cross-modal adaptive attention network is constructed to achieve effective fusion of text and multi-modal data. This network dynamically focuses on key information through an adaptive attention mechanism, automatically adjusting the weights of each modality to fully leverage the complementarity between modalities. In experiments, multiple representative social media datasets are selected for validation, and the results demonstrate that, compared with traditional methods, the proposed method significantly improves evaluation metrics such as accuracy and recall.

Query efficiency remains a primary challenge for black-box adversarial attacks, largely due to the difficulty of searching in high-dimensional input spaces. This paper introduces DSG-ALA, a framework that addresses this challenge through a tight coupling between saliency-induced spatial sparsity and the Artificial Lemming Algorithm (ALA). Our key innovation constructs a highly sparse perturbation subspace via a zero-order estimated, dual-scale saliency mask, directly constraining ALA’s search process to regions most influential to the model’s output. Comprehensive evaluations demonstrate that DSG-ALA reduces average queries by 10.4% on ImageNet/ResNet-50 while achieving superior visual quality compared to the strongest baseline. This work establishes spatially sparse, saliency-guided perturbation as an effective paradigm for developing query-efficient and imperceptible black-box attacks.

Abstract Encrypted web traffic and evolving Internet technologies pose an increasing challenge to network traffic analysis. However, existing traffic classification methods, though effective, require large labeled datasets and complex training. This makes sustaining them prohibitively expensive and difficult in real-world scenarios. To narrow this gap, we propose a novel training-free few-shot network traffic classification framework based on large language models (LLMs). By integrating meta-learning with LLMs, it reduces reliance on labeled data, eliminates task-specific training, and improves performance. Specifically, we first apply an efficient feature extraction method to extract features from traffic flows. We then design meta-tasks that combine task descriptions with textualized features to produce natural language meta-task formulations. Building on these meta-tasks, the LLM performs reasoning to carry out traffic classification. Finally, to mitigate hallucination in the LLM outputs, we exploit the temporal characteristics of network traffic and aggregate predictions over samples within a defined time window. Extensive experiments on three widely-used encrypted traffic datasets demonstrate that our proposed framework outperforms the state-of-the-art methods, achieving an average absolute improvement in F1 score of 9.75, 9.82, and 12.06 percentage points on the three datasets, respectively.

Abstract Accurate measurement of bolt preload is critical for ensuring the performance and operational safety of mechanical structural connections. However, traditional modal-conversion-based longitudinal and transverse wave testing methods suffer from low transverse wave excitation efficiency, insufficient signal stability, and limited feature-extraction accuracy. Taking an M36×3-350 mm bolt as the research subject, this study proposes an ultrasonic testing method based on direct excitation of transverse waves using the D15 shear mode of a piezoelectric transducer. It simultaneously incorporates the / time ratio feature to characterize the preload and employs a cubic spline interpolation algorithm to reconstruct the discretely sampled signals, thereby improving peak localization accuracy and enhancing the precision of time-of-flight extraction. The research results indicate: (1) Compared to traditional modal conversion excitation methods, direct excitation via the D15 shear mode significantly improves transverse wave excitation efficiency and modal purity. (2) The sound time ratio between transverse and longitudinal waves exhibits a significant linear negative correlation with bolt pretension and can serve as an effective feature parameter for characterizing changes in bolt pretension; numerical simulations agree well with experimental results, with the overall measurement error controlled within 3%. (3) The use of a cubic spline interpolation algorithm enables the raw sampling interval to be increased from 0.01 μs to 0.0001 μs, improving the positioning accuracy of the first-arrival time of ultrasonic waves and enabling the effective identification of minute changes in preload. These research findings provide a theoretical foundation and technical support for bolt connection condition monitoring and intelligent structural health monitoring under complex operating conditions.

Efficient fault detection in photovoltaic (PV) modules is pivotal for maintaining the operational safety and stability of solar power generation systems. Despite significant advancements in deep learning-based fault detection, existing models often struggle with limited feature extraction and low efficiency, impeding real-time deployment in complex environments. To address these issues, this paper proposes a lightweight and highprecision fault detection model based on YOLOv7-tiny. The proposed model integrates a thermal infrared and temperature information fusion module into the backbone network to enhance multi-modal feature representation. To ensure computational efficiency, GhostConv is utilized to replace standard convolutional layers, minimizing parameter overhead without compromising performance. Furthermore, a cross-scale lightweight attention module is designed to improve the contextual awareness and localization performance for small targets. Prior to the detection head, a position information perception module is introduced to enhance global semantic modeling and local directional sensitivity. Experimental results demonstrate that the model achieves mAP50 and mAP50:95 of 93.5% and 59.4%, respectively, on an infrared photovoltaic defect dataset. Combining a lightweight architecture with improved accuracy and robustness, the proposed model demonstrates significant potential for integration into intelligent photovoltaic operation and maintenance systems.

By applying nonlinear vibration-to-sound modulation technology to bolt loosening detection, this paper proposes a new experimental setup and signal-processing method. A linear swept-frequency signal is used to excite low-frequency vibrations, while a fixed-frequency sine wave is used for high-frequency ultrasonic excitation. First, a fractional Fourier transform is applied to the vibration-acoustic modulation signal to transform it into the optimal fractional domain where the energy of the swept-frequency signal is concentrated; next, the swept-frequency signal undergoes a masking filter, and the filtered signal is transformed back to the time domain; finally, the time-domain signal is transformed back to the frequency domain, and the amplitudes of the sum and difference frequency components of the high-frequency signal are extracted as damage parameters. The effectiveness of this method in bolt loosening detection was verified through bolt connection tests, with the applied tightening torque ranging from 10 to 30 N & centerdot;m. This method is simple to operate and highly robust, making it a reliable approach for detecting the degree of loosening in bolt connections.

Internet traffic classification has attracted significant attention due to its importance in traffic engineering and network security. Although deep learning–based approaches have demonstrated strong effectiveness for encrypted traffic classification, they typically rely on large volumes of labeled data, which are costly and difficult to obtain in real-world settings.To narrow this gap, we propose FlowMatch, a semi-supervised framework for encrypted network traffic classification. Specifically, we segment traffic by flows and transform each flow into a two-dimensional grayscale image representation. Subsequently, we devise a set of task-specific augmentation strategies for traffic images. Building on this, we adopt a semi-supervised training paradigm that combines consistency regularization with pseudo-labeling, enforcing prediction consistency between weakly and strongly augmented samples to fully exploit unlabeled data and enhance the model’s generalization ability. Extensive experiments on three widely used encrypted traffic datasets show that our framework outperforms competitive baselines, achieving F1 scores of 82.29, 96.09, and 90.23 with only 10 labeled samples per class. Additional ablation studies further validate the effectiveness of each component.

Deep neural networks exhibit critical vulnerabilities to adversarial examples, posing severe threats to security-sensitive applications. In practical black-box settings where attackers only have query access, existing methods suffer from fundamental limitations: significant query consumption, persistent over-optimization, and inefficient exploration-exploitation balance. To address these limitations, we propose AdaAttack, a novel gradient-free adversarial attack framework that employs an energy-guided strategy coordinator, dynamically integrating four adaptive strategies for efficient boundary exploration and refinement. A second key component is a query-aware saturated margin loss that prevents over-optimization through confidence-based early stopping. Extensive evaluations on MNIST, CIFAR-10, and ImageNet demonstrate that AdaAttack achieves high attack success rates (92.5–100%) with significantly fewer queries compared to state-of-the-art methods. While requiring more queries on MNIST, it maintains perfect attack success, highlighting its suitability for complex, high-dimensional domains. These results confirm that AdaAttack effectively resolves the core trade-offs in black-box attacks, offering a robust and efficient solution for real-world deployment.

This paper studies the effect of buy-online-and-pick-up-in-store (BOPS) implementation for an omnichannel retailer providing online and offline channels, where product price and free-shipping threshold are decision variables. Especially, for the online channel, the retailer employs a contingent free shipping (CFS) policy. Based on the different characteristics of purchasing channels, we first construct the utility functions of purchasing channels and then employ the principle of maximum utility to obtain demand functions. Further, we develop optimization problem models with and without BOPS channels to study the effect of BOPS. Then, based on the original problem, we consider two extension models (i.e., the difficulty of finding suitable add-on product and the market competition). Finally, we conduct some numerical experiments to obtain some managerial insights about the BOPS implementation. Our results show that the BOPS implementation can increase the total demand. However, the BOPS channel is beneficial for omnichannel retailers only under the small BOPS hassle cost and the lower BOPS operation cost. Besides, when the CFS policy has less attractiveness to online customers, the BOPS implementation is more beneficial to omnichannel retailers.

Infrared small target detection faces persistent challenges due to low signal-to-noise ratio and the scarcity of texture information in complex backgrounds. Existing deep learning methods are often limited by the "vanishing target" phenomenon in deep layers and the reintroduction of background noise during feature fusion. To address these issues, this article proposes a novel dual-stream cascade frequency decoupling and difference enhancement transformer (DCFD-DETR). This framework introduces a parallel dual-stream architecture that deeply integrates frequency-domain physical priors with spatial features through cross-domain semantic interaction and complementary fusion. Specifically, a wavelet-based frequency-spatial decoupling module and a spatial purification module are constructed in a cascade manner to explicitly decouple high-frequency target singularities from low-frequency background clutter, utilizing the latter to effectively suppress background interference in the spatial domain. To facilitate efficient cross-domain interaction, a cross-guided hybrid attention fusion block is designed to align semantic features. Furthermore, a difference-enhanced semantic-gated fusion module is proposed, employing a "Difference is the Target" mechanism to enhance target salience while filtering out consistent noise. Finally, to alleviate the optimization bottleneck of small target regression, a dynamic auxiliary-box optimization strategy is introduced. By establishing a dual-head parallel mechanism with shared query constraints, this strategy accelerates convergence and ensures precise pixel-level localization. Extensive experiments on the NUDT-SIRST and infrared small target detection-1K benchmarks demonstrate that DCFD-DETR outperforms existing state-of-the-art methods, exhibiting superior robustness and detection accuracy in complex scenarios.

Remote sensing imagery presents unique challenges for object detection due to wide fields of view, complex backgrounds, and the dense distribution of small targets, often rendering traditional methods ineffective. To address these limitations, we introduce GSS-YOLO, a lightweight network tailored for remote sensing environments. Our architecture integrates a Spatial Information Aggregation (SIA) module within a Cross-Stage Partial Network (C3) to optimize both detection accuracy and processing efficiency. Furthermore, we incorporate Spatial Pyramid Dilated Convolution (SPD-Conv) to enhance adaptability to low-resolution inputs, and embed a Global Context-Aware Module (GCAM) prior to the detection head to refine multi-scale feature representation. Evaluations on the USOD, VisDrone2019 and DIOR datasets demonstrate that GSS-YOLO achieves superior precision, recall, and robustness across both color and grayscale imagery, all while maintaining a lightweight architecture. Validated by ablation studies, this approach provides an efficient and robust solution for small target detection in complex remote sensing scenarios.

Accurate segmentation and quantification of the eyeball and lens from MRI images are crucial for clinical diagnosis and treatment planning of ocular diseases. Traditional methods for analyzing eye structures in MRI have drawbacks including low segmentation accuracy and reliance on laborious, time-consuming manual processes. To solve these problems, we propose a SEDP-SegResnet model for segmentation of the eyeball and lens structures from 3D MRI images. The framework takes SegResnet as its backbone network and incorporates a 3D-SE layer to handle deep features from decoder, 3D-SE layer assigns different weight information to the feature map channels through squeeze and excitation mechanism. Moreover, skip connections in the U-shaped architecture model are replaced with Dynamic Deep Feature Prefusion (DDFP) modules. The DDFP can achieve in-depth fusion of encoder and decoder features based on global information, thereby enhancing 3D image context comprehension of the model. The performance of SEDP-SegResnet is evaluated through a series of experiments using a proprietary dataset of orbital MRI scans. The results show that SEDP-SegResnet outperforms current mainstream 3D deep-learning-based segmentation models across multiple evaluation metrics including the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU). The model achieves robust performances in segmenting margin of eyeballs and blur-edge lenses. SEDP-SegResnet achieves a DSC of 96.81% for eyeball segmentation and 90.57% for lens segmentation, superior than a variety of commonly used segmentation models. It provides a more accurate, automated and robust method for the segmentation and quantification of eyeball and lens in MRI, offering an advanced computer-aided diagnosis tool.

This paper addresses the capacitated vehicle routing problem for omnichannel retailing with multiple types of time windows (hard time windows and soft time windows) and multiple types of products simultaneously. The problem aims to transfer multiple products from a central warehouse to stores and satellites, where split delivery is allowed since the demands of stores or satellites are large and the different product types have different volumes. Based on different characteristics of purchasing channels, the time window of a store is a hard constraint while the time window of a satellite is a soft one. This type of vehicle routing problem exists extensively in omnichannel retail distribution systems in the logistics industry of China, which considers multiple purchasing channels for customers. Since this type of vehicle routing problem is an NP-hard problem and more complicated than the conventional vehicle routing problem with time windows, we design an adaptive large neighborhood search (ALNS) method to solve it. Finally, numerical experiments are conducted to evaluate the effectiveness of the proposed algorithm and reveal some managerial insights. The numerical experiments show that the proposed algorithm can obtain a high-quality solution under a shorter computation time compared with the state-of-the-art MIP solver (Gurobi).

In the field of cybersecurity, the deluge of alerts presents a significant challenge to human review capabilities. Despite existing solutions, there is still an urgent need for more advanced methods to improve the effectiveness and accuracy of false alerts filtering. In this paper, we propose FAF-BM, a cutting-edge approach that integrates the BERT model, semi-supervised learning and active learning to enhance alert filtering capabilities. FAF-BM leverages the fine-tuned BERT model to fully exploit the deep semantics of alerts without being constrained by the format of the alerts. Subsequently, the semi-supervised learning is dedicated to mining the hidden potential within unlabeled data, thus expanding the learning scope beyond the confines of labeled datasets. In addition, the active learning strategically utilizes the expertise of security professionals to guide the learning process, ensuring that the approach adapts to the evolving threat landscape. Through a series of experiments, it has been demonstrated that FAF-BM not only improves the effectiveness of the filter but also enhances the generalization ability of dealing with the heterogeneity of alerts.

Addressing the real-time performance and robustness challenges of Unmanned Underwater Vehicle (UUV) systems operating in dynamic and complex environments, this paper presents a distributed heterogeneous multi-UUV dynamic task allocation algorithm, Dynamic Consensus Based Bundle Algorithm (D-CBBA), built upon an enhanced Consensus Based Bundle Algorithm (CBBA). First, we formulate a heterogeneous multi-UUV dynamic task allocation optimization model that incorporates UUV capabilities, task types, and time-window constraints. Next, we design a dynamic task management mechanism that supports time-triggered task insertion and state-based task cancellation logic, using task completion detection to eliminate invalid allocations. We further optimize the compatibility matrix and the bidding score function by taking into account differences in task types and UUV capabilities. Finally, we introduce a fault-tolerant mechanism for UUV failures: a time-shift compensation algorithm distinguishes between completed and pending tasks, enabling dynamic task inheritance for failed UUVs and rapid system reconfiguration. The simulation results demonstrate that the proposed method effectively solves heterogeneous multi-UUV task allocation problems in complex and dynamic environments. Compared with the baseline CBBA, the improved algorithm achieves higher task completion rates, resolves task conflicts in complex dynamic scenarios, reduces computational resource waste, and accelerates convergence.

With the continuous progress of remote sensing technology, the demand for means of detecting small targets in remote sensing images is escalating. The significance of detecting small targets in remote sensing images lies in enhancing the ability to identify small and elusive targets and the detection accuracy against complex backgrounds, holding significant application value in military reconnaissance, environmental monitoring, and disaster early-warning systems. Firstly, the minuteness of certain targets in relation to the entire image in which they occur, particularly when the camera is situated at a higher altitude, renders them difficult to detect. Secondly, the varying background and lighting conditions in remote sensing images further complicate the detection task. Conventional target detection methods are frequently incapable of addressing these complexities, resulting in a reduction in detection accuracy and an increase in false alarms. Hence, in this paper, we propose a lightweight remote-sensing image target detection network model, CBGS-YOLO, created by introducing the Ghost module to decrease the model parameters, applying the SPD-Conv module to optimize downsampling, and integrating the convolutional block attention module to enhance detection accuracy. The experimental outcomes demonstrate that CBGS-YOLO outperforms other models when applied to the DB_Licenta and USOD datasets, significantly enhancing detection performance for small targets. Compared with YOLOv9, this model can reduce the number of parameters from 7.10 M to 5.12 M, and the average precision (mAP) is effectively improved. The model strengthens the ability to identify small targets against complex backgrounds while maintaining lightweight properties and possesses remarkable application prospects and practical value.

Aiming at the problems of complex sensor layout, low detection accuracy when using guided wave to detect the elbow's erosion degree, this research presents a new approach based on asymmetric guided wave and Genetic Algorithm-Least Support Vector Machine (GA-LSSVM). The program for calculating elbow's dispersion curve was compiled using the semi-analytical finite element (SAFE) method. Through numerical simulation, it is found that the bending mode in the guided wave signal at the outer arch back position of the receiving end decreases with the increase of the erosion degree, which leads to the decrease of the signal's time domain energy. Finally, The GA-LSSVM model were used to predict various elbow specimens' residual wall thickness, and the results shows the high prediction accuracy. The average accuracy of predicting the new elbows is 87.63%. This research provides a feasible scheme for detecting elbow's erosion degrees in engineering.

Long-wavelength (LW) cryogenic (from 80 down to 40K) infrared focal plane arrays (IRFPAs) are extensively employed because of their sensitivity and rapid response. Readout circuits (ROIC) with high dynamic range (HDR) can significantly improve the signal-to-noise ratio (SNR) of LW-IRFPAs due to enhancement of well capacity. However, neither conventional current-voltage nor current-frequency modulation can avoid analog circuits, making the ROIC susceptible to interference [1]–[8]. In these ROICs, active analog circuits are usually the primary power consumption item. Moreover, the long data transfer bus with a high swing across the pixels consumes power and introduces crosstalk. To address these issues, an LW-IRFPA architecture with pixel-paralleled light-driven current-to-phase A/D conversion (IP-ADC) is implemented in this work. The key circuit includes: 1) a light-current-controlled oscillator (LCO) based modulator for IP-ADC that is powered only by photocurrent without extra power supply; 2) a phase-reconstruction circuit to enhance low-light performance while reducing quantization noise by sub-phase extraction; 3) a pixel-level 20b asynchronous gray-code counter for high equivalent full well capacity during phase quantization; and 4) a low-swing bus data transfer circuit to reduce transmission power consumption and crosstalk. Overall, this work demonstrates a $320\times 256,\ 30\mu \mathrm{m}$ pixel pitch, $10.55\mu\mathrm{m}$ wavelength quantum-well infrared photodetector (QWIP) IRFPA. The ROIC used in IRFPA is fabricated in a $0.18\mu \mathrm{m}$ standard CMOS process. The performance of this IRFPA achieves 120.4dB HDR, 2.2mK noise equivalent temperature difference (NETD) within a typical 6.9mW power consumption.

The Provenance Graph has been widely applied in host anomaly detection due to its ability to detect unforeseen threats, reveal correlations between host behaviors, and the superior performance offered by Graph Neural Networks. However, a barrier to the practical adoption of Provenance-based intrusion detection systems is the lack of interpretation. Unfortunately, existing interpretation approaches are proposed for either non-security domains or other kinds of anomaly detections like tabular, homogeneous graph or network intrusion detections, which are not suitable for Provenance-based intrusion detection systems and fail to satisfy special requirements in security domains. In this paper, we propose PGAid, a method aiming to interpret Provenance-based intrusion detection in security domains. Firstly, PGAid addresses the different interaction features between nodes in heterogeneous graphs by layering the provenance graph to be explained according to the types of source and destination nodes, segmenting it into different subgraphs for explanation. Then, we search for reference edge groups within each subgraph, adjusting the priority during the search process. Finally, we verified the effectiveness of our method from fidelity and stability, which respectively surpassed other baselines by 10% and 2-5%.