Accurate high-resolution spatio-temporal air quality inference based on sparse monitoring stations is essential for environmental governance and public health. However, prevailing deep learning approaches often treat inference as direct regression on observed air quality concentrations, neglecting the distinct effects of a long-term static baseline and short-term dynamic perturbations on the inference. This limits the model's generalization ability and may cause spatial bias in the inference. In this paper, we propose a novel Spatio-Temporal Air Quality Inference Network (SAQIN) model, which provides an explicit decomposition of the static and dynamic effects in the inference. The static branch integrates high-resolution semantic segmentation, global context from remote sensing, population density, and elevation to construct a high-fidelity environmental prior. The dynamic branch employs multi-head self-attention mechanism to jointly encode geographic relationships and inter-pollutant chemical coupling effects. We further develop a residual learning module to model the inference process as a residual correction anchored to multiple reference stations and aggregates predictions through stepwise multi-station fusion with distance-weighted averaging, which successfully eliminates the Voronoi-partition-induced discontinuities and yielding a globally smooth physically plausible air quality inference field. SAQIN has been shown to demonstrate superior accuracy and strong zero-shot cross-domain generalisation when evaluated on large-scale real-world datasets, thus outperforming state-of-the-art methods across a range of criteria pollutants.
Rationalization models have recently garnered significant attention for enhancing the interpretability of natural language processing by first using a generator to select the most relevant pieces from the text with respect to the label, before passing the text input to the predictor. However, the robustness of the rationalization models is not sufficiently investigated. Specifically, this paper explores the robustness of rationalization models against backdoor attacks, which has been ignored by previous studies. Surprisingly, we find that conventional backdoor attack techniques fail to inject triggers into the rationalization model because its generator can filter out bad triggers. Considering this, we further propose a novel backdoor attack method named as BadRNL designed specially for the rationalization models. The core idea of BadRNL is first to search for the personalized trigger for each specific dataset and then manipulate the rationales and labels to conduct attacks. Besides, BadRNL controls the order of sample learning through poison-priority sampling strategies. Experimental results across five diverse NLP datasets-FEVER, MultiRC, Beer, Hotel, and Movie-demonstrate that BadRNL consistently achieves near 100% attack success rates (ASR). Crucially, the method maintains high classification accuracy and rationale quality on benign samples, highlighting a significant security risk for interpretable NLP systems deployed in real-world supply-chain scenarios.
Multiple feature fusion networks integrate diverse features to enhance predictive performance, yet they face two key challenges that require further exploration. First, most studies focus on feature fusion within known closed-set classes, overlooking the more practically significant open-set scenario with unknown classes. Second, these approaches often encounter adaptability conflicts, including expressivity limitations from inaccurate patterns captured in handcrafted features, overadapting due to excessive adaptation to the training data in deep features, and inconsistent feature hierarchy challenges in mixed-feature scenarios. To address these challenges, we propose OpenEnhancer to improve open-set learning through multiple feature fusion. We introduce a pseudo-unknown mechanism and a feature-specific sparse learning layer to form an enhanced feature-specific sparse module that better adapts the model to open-set scenarios and mitigates adaptability conflicts like expressivity limitations and overadapting. The processed features are integrated using a multi-feature attention fusion module, which further alleviates the challenges posed by inconsistent feature hierarchies. Finally, we employ an open-set loss regime to reduce spurious confidence in known classes when the model is exposed to the unknown, thereby improving recognition accuracy. Open-set simulation experiments on multi-feature datasets demonstrate that OpenEnhancer not only achieves strong recognition performance but also effectively mitigates adaptability conflicts. The datasets and source code are publicly available to ensure reproducibility: https://github.com/dushide/OpenEnhancer.
Most existing ictal stereoelectroencephalography (SEEG)-based seizure onset zone (SOZ) localization methods rely on patient-specific training, limiting their clinical applicability due to the scarcity of seizure recordings and substantial inter-patient variability. Consequently, robust patient-independent SOZ localization remains a major challenge. In this work, we propose a deep learning approach for patient-independent SOZ localization using ictal SEEG recordings, aiming to improve cross-patient generalization while preserving seizure-related temporal characteristics. To mitigate domain shifts across subjects, we introduce a clinically guided feature learning strategy that combines a cross-frequency coupling (CFC) mechanism to capture SOZ-related abnormal interactions across frequency bands with a self-comparison (SC) mechanism to emphasize seizure-onset evolution patterns within SEEG channels. We further incorporate seizure detection as an auxiliary task within a multi-task learning framework to provide seizure-onset-related temporal supervision, thereby improving the temporal awareness and generalizability of SOZ localization. Experiments on the public OpenNeuro HUP dataset demonstrate substantial improvements over existing methods, while additional evaluations on a private clinical dataset further validate the robustness and cross-patient generalization capability of the proposed method. Moreover, comparisons between the learned CFC representations and clinically established phase-amplitude coupling (PAC) metrics reveal consistent physiological patterns, supporting the interpretability of the learned representations.
This paper addresses the event-triggered impulsive control problem for nonlinear switched systems with time delay. The main challenge is that the delayed state may originate from a historical interval governed by a subsystem different from the current one, while switching may also occur before the next impulse is generated. To address this issue, a mode-dependent event-triggered mechanism is developed by incorporating an exponential attenuation term associated with the initial historical information. Based on the Lyapunov-Razumikhin method and the average dwell time technique, sufficient criteria are established to ensure exponential stability of the considered system and rule out Zeno behavior. An explicit lower bound on the inter-event times is also obtained, which characterizes the admissible triggering intervals under the combined influence of switching and delay. Since the impulse input is applied only at triggering instants, no continuous control signal needs to be transmitted between two successive triggers, and the communication burden associated with control transmission from the controller to the actuator can be reduced. The proposed theoretical results are further applied to the synchronization of switched neural networks, where the design conditions for the triggering mechanism and impulsive control gain are formulated as linear matrix inequalities. Numerical simulations and an image encryption application demonstrate the effectiveness and practical applicability of the developed method.
Recently, the Mamba architecture has demonstrated outstanding performance in natural language processing and general vision tasks, and has been rapidly extended to the field of Remote Sensing Change Detection (RSCD). However, existing Mamba-based change detection methods lack explicit modeling for precise difference features. The process of extracting difference features is prone to amplifying irrelevant noise, making it challenging to extract subtle changes and boundary information in complex remote sensing scenes. To address these challenges, we propose a novel network, DFMamba, integrating two innovative modules. First, the Difference Boost Mamba Fusion (DBMF) module employs differential contrast features to enhance the differences in the original features, thereby further amplifying the finer distinctions between the bi-temporal images. Second, the Adaptive Frequency-Aware (AFA) module generates adaptive frequency-domain signals using learnable filter coefficients, which suppress the irrelevant high-frequency noise introduced by the differential feature enhancement while enhancing information about changing boundaries. To evaluate model performance, experiments are conducted on five benchmark datasets (LEVIR-CD, LEVIR-CD+, WHU-CD, SYSU, and CDD). Results demonstrate that DFMamba achieves competitiveness comparable to multiple state-of-the-art models across all datasets, validating its effectiveness. The source code of our proposed model will be released at https://github.com/PriscaCommit/DFMamba.
RGBT object tracking takes advantage of the complementary properties of RGB and thermal infrared (TIR) modalities. However, many existing methods focus on fusion within a single domain, either spatial or frequency, without fully exploiting the complementarity of multiple domains. This limits the interaction between domains and makes trackers less robust under severe conditions. To improve feature representation by leveraging both spatial and frequency information, we propose a Frequency domain information-Aided Spatial domain Feature Fusion framework for RGBT tracking (FASFF), which mainly consists of the Frequency domain Attention Enhancement Module (FAEM) and the Spatial-Frequency feature interaction Fusion Module (SFFM). The FAEM employs the Discrete Cosine Transform Attention block (DCTAttn), which highlights informative frequency components while integrating spatial and frequency domains information, and the Cross-modality Collaborative Attention block (CCAttn), which suppresses noise via cross-modal feature refinement to enhance intra-modality feature representation. The SFFM further harnesses the interaction between the spatial and frequency domains to collaboratively refine feature representation, leveraging complementary information from the frequency domain to enhance spatial domain representations. We evaluate the FASFF on three benchmark datasets: RGBT210, RGBT234, and LasHeR. The experimental results show that the FASFF maintains stable and reliable performance across different tracking scenarios, verifying its effectiveness for RGBT object tracking.
Class imbalance presents a persistent bottleneck in machine learning, biasing models toward majority classes and degrading performance on rare events. Active learning partially alleviates this by focusing annotation efforts on underrepresented samples. However, this issue becomes more pronounced in multi-domain learning settings, where both intra-domain imbalances and shifts across domains complicate model training. Current multi-domain active learning techniques typically neglect the interaction of these dual imbalances, resulting in biased selection strategies and limited ability to generalize across domains. To tackle this, we introduce a comprehensive framework featuring a dynamic category-scaling weight mechanism that hierarchically addresses imbalance at the sample and domain levels. This mechanism assesses the relative difficulty of each class to guide effective sampling, while concurrently constructing a progressively refined "Ideal Domain" through the adaptive integration of well-performing domains. This ideal representation further supports a minimax optimization process to allocate annotation budgets both fairly and robustly. Extensive evaluations on various standard datasets confirm that our approach consistently surpasses leading methods in managing complex imbalanced scenarios across multiple domains.
Implicit Q-learning, a prominent approach within the implicit value learning (IVL) paradigm, mitigates distribution shift in offline reinforcement learning (offline RL) by avoiding out-of-distribution action queries. However, current IVL methods lack stepwise training analysis in value learning and commonly assume the behavior policy to be optimal, both of which overlook more realistic suboptimal scenarios. This work presents cautious implicit value learning (CIVL), a novel offline RL method that redefines the advantage weight to identify a previously unrecognized constraint in IVL training. By establishing a formal connection between the identified constraint and the target V-network, CIVL provides theoretical motivation for introducing the target V-network. In policy learning, CIVL introduces a negative-action penalty (NAP) that penalizes actions with negative advantage values, thereby better handling suboptimal actions. Since the NAP can be shown to stochastically induce an approximate gradient penalty on the Actor, CIVL introduces a floating term that synergistically increases the probability of the gradient penalty occurring. Finally, CIVL further incorporates trajectory-position weighting to provide temporally fine-grained control over the V-network's bias. Benchmark experiments demonstrate that CIVL outperforms existing IVL methods in terms of normalized score and training stability.
Current vision-language multimodal long-term tracking methods are highly dependent on large-scale data training and complex cross-modal models. This not only leads to substantial computational overhead but also restricts their deployment and application in resource-constrained scenarios. To resolve this contradiction, we propose a training-free, fast-response post-processing tracking framework named VL-METrack, which dynamically fuses motion estimation and localization networks. This framework transforms the long-term tracking task from a data-driven model training problem into a heuristic decision-driven online adaptive decision-making problem. Specifically, we first adopt a general pre-trained network to provide initial target localization and confidence assessment. When the confidence indicates tracking ambiguity, instead of relying on deeper networks, our method leverages a collaborative processing mechanism consisting of optical flow and correlation filters. This mechanism dynamically fuses optical flow and the localization network through confidence based adaptive weighting, while incorporating an error correction mechanism that judges the rate of change in target orientation. Subsequently, we employ a dynamic template update strategy and establish a global re-detection mechanism, maintaining template timeliness through an adaptive update process and determining the optimal position via multi-template comparison within an expanded search region. Experimental results on the OTB, LaSOT, and TNL2K datasets demonstrate that our method achieves performance comparable to state-of-the-art approaches while maintaining low computational complexity and reduced resource consumption. The code is available at: https://github.com/yll1214/VL-METrack.
The decomposing problem of collective mixture encompasses various concrete and typical examples, such as mixture models (e.g., GMM), clustering, and HMM, etc. These problems commonly involve latent variables related to sample classification, which is their fundamental characteristic. To address this, researchers typically use the variational inference method (specifically, EM algorithm) to solve the problem. However, this method requires prior information about the cluster number, which can sometimes be difficult to obtain. In this paper, we propose an Extended Variational Mode Decomposition (EVMD) method and its serialized form (SEVMD) to tackle the decomposition issue on collective mixture. These two algorithms correspond to two different prior conditions of the modal number. They are both variational methods and from the perspective of mode decomposition. Our experimental results demonstrate that the EVMD method can effectively recover the original components when the modal number is known. On the other hand, SEVMD successfully extracts the components in a serial manner even when such information is not available.
With the wide application of deep neural networks, the research on adversarial robustness has become one of hotspots, especially the issue of fairness in adversarial robustness has garnered the wide attention. Significant disparities in adversarial robustness across different classes could degrade the overall performance, because the effectiveness of the defense mechanism is constrained by the class with the lowest robustness called the worst class, that is the phenomenon known as the "wooden barrel effect". Most of the existing robust fairness algorithms were designed based on specific experimental metrics, neglecting analyzing the intrinsic causes of the robust fairness. And the trade-off of robust contribution between easy and hard classes was not considered. In addition, the robust balance between hard classes and their similar classes was ignored too. The above issues actually limited the further improvement of robust fairness. Therefore, A Frequency-Aware Adversarial Robust Fairness Algorithm (FAFA for short), which explores the inherent causes of fairness issues from frequency perspective, is proposed in the paper. And a Frequency-Aware Fair Adversarial Training method called FAAT based on the contribution of frequency components is introduced, and different loss weights and perturbation constraints are assigned to various classes to enhance the robustness of hard classes. Additionally, a Frequency-Aware Fair Fine-tuning Algorithm (FAFT) which redistributes features between hard classes and their similar classes is implemented, achieving a more balanced robustness distribution and improve the performance of the worst class. Extensive experiments were conducted on the CIFAR-10, CIFAR-100, and STL-10 datasets, and the results demonstrate that our proposed method significantly enhances the robustness of the worst class, the accuracy of the worst class reaches 34.76% using the ResNet18 model on the CIFAR-10 dataset under A3 attack, with an average improvement of approximately 5% compared to existing methods. Meanwhile, the overall robustness reaches 48.47%, and the comprehensive metric ρrob based on TRADES reaches 0.55.
As a foundational technology in autonomous driving, 3D object detection plays a critical role by enabling vehicles to perceive their surroundings. However, LiDAR suffers from severe point cloud degradation in adverse weather, while 4D Radar, though more robust, introduces low-resolution and noisy measurements. Fusing LiDAR and 4D Radar is expected to enhance robustness, yet existing methods lack the ability to dynamically adapt to weather-induced sensor degradation and remain vulnerable to radar noise. To address these issues, we propose DDMDGF, a novel LiDAR-4D Radar fusion framework for weather-robust 3D object detection. The framework introduces a Semantic-guided Foreground-aware Denoising (SFD) module, which leverages enhanced semantic feature prediction and hybrid dynamic thresholding to suppress 4D radar noise while preserving critical foreground structures. In addition, we design a Multi-scale Dual-attention Gated Fusion (MDGF) module, which employs parallel intra-modal and inter-modal backbones to extract complementary features, and uses a channel-spatial joint attention gate to adaptively balance modality contributions under varying weather conditions. Extensive experiments on the VoD and K-Radar datasets demonstrate the effectiveness of the proposed method: DDMDGF achieves +7.3% AP3D and +4.9% APBEV improvements over L4DR on K-Radar, and under the most severe fog level in VoD-Fog, it outperforms L4DR by +1.6%, +1.8%, and +1.5% mAP on cars, pedestrians, and cyclists, respectively.
Emotion recognition based on electroencephalogram (EEG) signals has gained interest owing to its potential in affective computing. In practice, EEG signal acquisition is highly prone to noise interference from electrooculography (EOG) and electromyography (EMG) signals, severely impacting the cross-subject emotion recognition performance of models in real-world scenarios. Although learning-based denoising approaches have been developed to address this issue, they rely on standalone autoencoders for reconstruction, resulting in weak integration with downstream tasks. This study proposes a novel global-regional anti-noise network (GRANet) that is specifically designed for robust cross-subject emotion recognition from noisy EEG data. During modeling, GRANet adopts a novel noising training strategy and similarity-based noise-robust (NR) loss function. According to the noising training strategy, GRANet can extract anti-noise features from noisy EEG data using the NR loss. Moreover, GRANet comprises a three-dimensional global convolutional neural network (3D-GCNN) and a three-dimensional regional convolutional neural network (3D-RCNN). These networks aim to separately extract global and regional anti-noise features from noisy EEG data by applying the noising training strategy. We validate the effectiveness of GRANet by comparing it with state-of-the-art (SOTA) cross-subject methods on four benchmark datasets (SEED, SEED-IV, SEED-V, and DEAP). Findings demonstrate that GRANet has superior performance on these datasets, particularly improving average accuracy by 20.97% on SEED-V from the best baseline. In a nutshell, GRANet can robustly handle real-world EEG noise challenges while maintaining high recognition accuracy, establishing it as a promising solution for practical emotion recognition applications.
The proliferation of dynamic graph-structured data necessitates Continual Graph Learning (CGL) to enable models to learn incrementally and retain past knowledge. Class-incremental learning (class-IL) in CGL is particularly challenging due to catastrophic forgetting. Existing CGL strategies, including recent prompt-based learning methods, often grapple with significant memory overhead and suboptimal alignment between pre-training objectives and downstream continual tasks. We propose TriP, a triple-prompt framework aligning pre-training and class-incremental objectives in continual graph learning. As a lightweight parameter isolation-based method, TriP utilizes feature-level and class-level prompts to precisely capture task-specific knowledge. It introduces a unified prompt template that aligns downstream classification with self-supervised pre-training objectives, which can bridge the semantic gap and maximize the generalization capabilities of pre-trained graph models. Extensive experiments on four public benchmark datasets demonstrate that our method significantly outperforms state-of-the-art approaches, illustrating its effectiveness in mitigating catastrophic forgetting and improving the efficiency of continual learning on graphs.
The power-law relationship between large language model (LLM) performance and the scale of pre-training corpora has driven remarkable advances, while raising serious ethical and legal concerns. Pre-training corpora often contain sensitive personal information, copyrighted content, or benchmark test data, and the opacity of pre-training corpora further exacerbates these concerns. Consequently, detecting pre-training data has become a significant research challenge. Most existing detection methods rely on intermediate outputs of LLMs (activations, token probabilities, or model loss), which are inaccessible in commercial LLMs, where only final outputs are available. Only a few early-stage attempts rely solely on final outputs by measuring lexical or semantic similarity between reproduced and original texts at the sentence or token level. Yet, these attempts overlook the dynamic nature of reproducibility: reproducibility fluctuates differently across reproduction positions, contextual spans, and expressions, between seen and unseen texts. This oversight narrows the gap of reproduction similarity, ultimately degrading detection performance. To address this limitation, we propose RADS-PDD (reproducibility-aware dynamic similarity–based pre-training data detection), which shifts black-box pre-training data detection from static similarity comparison to reproducibility-aware dynamic similarity modeling. RADS-PDD incorporates three reproducibility-aware mechanisms, positional gain weight, continual gain weight, and triplet occurrence probability, to quantify varying reproducibility by dynamic similarity, amplifying the reproduction similarity gap between seen and unseen texts. Extensive experiments across two representative datasets, multiple data domains, and diverse LLMs demonstrate that RADS-PDD consistently outperforms detection methods that rely solely on final outputs and achieves performance comparable to detection methods that require intermediate outputs.
Continual multi-view clustering (CMVC) aims to discover cluster structures from multi-view data that arrive sequentially and exhibit distribution shifts. However, existing incremental anchor-based frameworks for CMVC still have some problems: 1. Anchors are learned in a purely reconstruction-driven way, resulting in ambiguous cluster semantics and higher sensitivity to noise. 2. Existing methods overlook the role of consistent anchor-level semantics, which limits their ability to suppress semantic drift. 3. It is tough to align the anchor graphs, weakening temporal consistency and degrading clustering stability. To address these limitations, we propose a novel continual multi-view clustering method with pseudo-label guided consistency anchor learning (CMVC-PCAL). In the initialization stage, we build a unified anchor graph representation by decomposing each anchor graph into a shared consensus component and a view-specific inconsistency component, and by introducing cluster-wise pseudo labels to align anchors with latent clusters, yielding a clearer and more noise-robust consensus anchor graph. In the continual stage, for each new data batch, a lightweight incremental scheme updates anchors and anchor graphs: an alignment matrix aligns historical and current anchors, and a consistency regularization term aligns their consensus parts, while a pseudo-label guided term jointly updates anchor pseudo labels and cluster centers to continuously refine anchor-level semantics. Finally, an alternating optimization algorithm is developed to efficiently train the model, and experiments on public multi-view benchmark datasets verify its effectiveness and efficiency in continual multi-view clustering scenarios.
It’s crucial for IoT network intrusion detection models to learn discriminative traffic representations from observable network traffic. Existing methods primarily rely on manually crafted features to represent network traffic. However, these manually constructed features depend heavily on the knowledge and experience of network experts and are difficult to generalize to diverse and constantly changing network environments. Therefore, leveraging a large amount of unlabeled raw network traffic to learn traffic representations from observable traffic patterns has become a research hotspot in the field of network intrusion detection. This paper proposes a new IoT network intrusion detection model called Bidirectional Encoder Representations from Transformer for Network Intrusion Detection (BERT4NID), which learns both intra-packet and inter-packet features by proposing two novel proxy tasks and pre-training on a large amount of unlabeled data to obtain traffic representations. The pre-trained model can be fine-tuned with a small amount of labeled data for downstream classification tasks and achieve excellent performance. With only 1% of labeled training data, the model’s F1 scores on MedBIoT and Kitsune datasets are 93.46% and 92.85%.
Deep multi-view clustering has been extensively studied due to its good ability to capture deep features and handle high-dimensional data. However, existing methods remain subject to the following shortcomings: (1) They typically employ single-layer fusion, which hinders the ability of fused features to comprehensively cover critical information; (2) They ignore the complex dependencies between views, thereby constraining the expressiveness of local features. We address these problems by proposing MambaMVC, a powerful state space model enhanced representation model for Multi-view Clustering. Specifically, we design dual-layer feature fusion to capture more comprehensive global information. First, we introduce a global structure-guided feature fusion learning module (GSFL) to conduct sample-level fusion as the first-layer stage. In the second-layer fusion, we design the global bi-directional Mamba module (GB-Mamba) to capture more intricate long-range dependencies between samples, which mitigate the information discrepancies between the preliminary global embedding feature and the view-specific embedding features. Meanwhile, we use the cross-view bi-directional Mamba module (CB-Mamba) to enhance information exchange between views, thus obtaining more comprehensive local information. Furthermore, contrastive learning is utilized to facilitate label consistency across cluster distribution, effectively reducing the influence of misleading information in specific views. Finally, we utilize the global target distribution to effectively guide the optimization of the cluster distribution in self-supervised learning. Experimental results on eight multi-view datasets demonstrate the effectiveness of the proposed model. Our code is publicly available at .
Underwater objects often suffer from severe visual degradation, including strong background noise induced by water impurities, blurred boundaries caused by optical scattering, and small object suppression in complex environments. These factors collectively pose significant challenges to underwater object detection (UOD). To address these issues, we propose the Multi-Scale Inverted Pyramid Network (MIP-Net) tailored for UOD. Moving beyond standard feature fusion paradigms, MIP-Net introduces explicit conceptual shifts through two key components: the Local Adaptive Contrast module (LAC) and the Multi-Scale Inverted Feature Pyramid Network (MSIFPN). Unlike conventional global attention mechanisms, LAC selectively calibrates intra-feature contrast layer-by-layer, establishing a theoretical basis for dynamic feature modulation that prevents high-frequency detail dilution. Simultaneously, MSIFPN transcends traditional FPN architectures by coupling a dual-pyramid bidirectional flow with a strict mathematical foreground-background separation strategy, effectively isolating target semantics from ambiguous water impurities. Experiments on the official DUO benchmark demonstrate that MIP-Net achieves an AP of 70.1%, surpassing state-of-the-art single-stage, two-stage, and Transformer-based methods, while maintaining a highly competitive trade-off between computational efficiency (Params/FLOPs) and accuracy. Furthermore, evaluations on the terrestrial COCO dataset yield an AP of 45.6%, confirming the strong cross-domain generalization capability of our framework. The code is publicly available at: https://github.com/YitengGuo/MIP-Net
