ANT-ET: An end-to-end multimodal framework for fine-grained encrypted traffic fingerprinting

With the prevalence of Internet of Things (IoT) technologies, the huge growth of IoT devices has also brought attention of cyber attackers. IoT botnets are rapidly spreading and evolving worldwide, causing serious risks to users and data. Machine learning (ML) has shown its effectiveness on threat detection. However, existing feature encoding and learning methods are unsuitable for resource constrained edge devices like the IoT gateway. In this paper, we propose a lightweight threat detection scheme called FlowSpotter. The flow imaging mechanism requires less feature extraction but preserves more spatial and temporal information. A lite convolution neural network architecture based on the state-of-the-art efficient building blocks is devised. For performance evaluation, we develop an IoT honeypot system that captures hundreds of thousands of IoT intrusions in the wild. Besides, FlowSpotter is implemented on Raspberry Pi for measuring the efficiency. Experimental results show that FlowSpotter not only outperforms 8 baseline models by achieving 99.8% accuracy and 0.07% false positive rate, but also consumes the least computing resources by taking less than 11 ms and 61 MiB memory for each detection.

Facial authentication on mobile devices has become prevalent in various applications. Face Liveness Detection, or Face Anti-Spoofing (FAS), focuses on identifying attempts by malicious users to impersonate someone else or hide their own identity. One specific branch within this field is active liveness detection, which involves analyzing both the input signal and user behavior while they perform a required task to verify the authenticity of the presented face. Despite the significant amount of research in FAS, active liveness detection remains mostly underexplored. This gap has led to outdated methods, insufficient testing of proposed active techniques in diverse scenarios, and a lack of comparative analysis between different approaches. In this paper, we explore these differences by comparing the performance of the latest existing close-up methods with baseline models using ResNet-18 and ResNet-50. Furthermore, we introduce a new model that builds on previous work, combining projective invariants with facial embedding for robust feature extraction. This approach directly improves upon existing techniques, surpassing other baselines in detecting spoofing attempts.

In order to execute cyber-security, intrusion detection systems (IDS) are developed to identify threats and irregularities in computer networks. An efficient data-driven intrusion detection system has been developed as a result of the use of artificial intelligence, particularly machine learning techniques. The proposed security model utilizes BGOTSVM to develop IDS systems starting from a security feature ranking process followed by model development using crucial features. The increasing sophistication of cyber threats necessitates robust and intelligent threat detection systems. This study uses the UNSW-NB15 dataset to demonstrate a Convolutional Neural Network (CNN)-based approach for financial fraud detection. To improve model performance, the suggested methodology incorporates data preparation techniques such as feature selection, one-hot encoding, and managing missing values. The CNN model, optimized through hyperparameter tuning, is compared against traditional machine learning (ML) models, including AdaBoost and naïve biased. Experimental findings show that CNN performs better than any baseline model, reaching the maximum accuracy (93.40%), precision (94.63%), recall (93.40%), and F1-score (92.81%). Performance evaluation metrics, classification reports, and confusion matrices further validate the CNN model’s ability to identify fraudulent activity. Comparative analysis indicates that deep learning techniques, particularly CNN, offer superior threat detection capabilities by effectively identifying complex trends in communication over network information.

Event-based cameras, inspired by biological vision, asynchronously capture per-pixel brightness changes, producing streams of events with higher temporal resolution, dynamic range, and lower latency than conventional cameras. These advantages make event cameras promising for human pose estimation in challenging scenarios, such as motion blur and low-light conditions. However, human pose estimation with event cameras is still in its early research stages. Among major challenges is information loss from stationary parts of the human body, where the stationary parts at instances cannot trigger events. This issue, inherent to the nature of event data, cannot be resolved in a short-range event stream alone. Therefore, incorporating motion cues from a longer temporal range offers a intuitive solution. This paper proposes a joint global and local temporal modeling network (JGLTM), designed to extract essential cues from a longer temporal range to complement and refine local features for more accurate current pose prediction. Unlike existing methods that rely only on short-range temporal correspondence, this approach expands the temporal perception field to effectively provide crucial contexts for the lost information of stationary body parts at the current time instance. Extensive experiments on public datasets and the dataset proposed in this paper demonstrate the effectiveness and superiority of the proposed approach in event-based human pose estimation across diverse scenarios.

Differentiable Neural Architecture Search has emerged as a powerful paradigm for automated network design, yet it suffers from a fundamental optimization inconsistency problem: Architectures optimized under continuous relaxation often fail to maintain their performance after discretization. To address this challenge, we propose Fisher-DARTS—a Fisher information-driven differentiable NAS framework. The proposed method introduces three key innovations: (1) a Fisher information-based momentum update mechanism that guides architectural parameters toward statistically significant operations, aligning the search objective with discrete deployment; (2) a progressive three-region pruning strategy that adaptively eliminates redundant operations with low Fisher information, ensuring architectural compactness; and (3) a cell-weighted fusion module that preserves multi-scale features across stacked cells. Additionally, the search space is expanded by incorporating attention mechanisms to enhance feature representation capability. The proposed framework is generic and applicable to a wide range of vision tasks. To validate its effectiveness, we apply it to gaze estimation—a core technology in multimodal human–computer interaction. Experimental results on three public datasets, MPIIFaceGaze, RT-GENE, and ETH-XGaze, demonstrate that Fisher-DARTS achieves mean angular errors of 3.22°, 5.45°, and 4.12°, respectively, outperforming hand-designed networks and existing NAS-based gaze estimation models. These results validate the effectiveness of the proposed Fisher-driven NAS framework and its generalization capability across diverse scenarios.

In this paper, a human-machine collaborative classification model based on Deep Forest (HMDF), is proposed for industrial product defect classification. Defect images are firstly classified by deep forest consisting of the multi-grained scanning module and the cascaded forest module. Based on the first-round result, difficult samples with low classification con-fidence are filtered out. A fraction of random selected difficult samples are manually identified and then used to classify the remaining difficult ones with the K-Nearest Neighbor (KNN) method with features extracted by the deep forest. The results of KNN are finally used to update the first-round classification results to correct errors in the initial results. The experimental results on the public data sets and the real-world data collected from the production line indicate that the proposed model could achieve superior performance compared with the baseline models in the industrial product defect classification task.

BDBRC: A Chinese military entity recognition model combining context contribution and residual dilatation convolutional networks

Simulating regional full-chain water balance using a novel surface water-groundwater coupled model

This paper addresses the problem of temporal moment localization with natural language, which aims to localize a target video moment according to a language description. A key challenge in this task is how to learn effective alignment between vision and language features extracted from an untrimmed video and a sentence description. Traditional methods generally utilize the vanilla attention mechanism to simulate soft alignment between extracted video and sentence features, however, they lack the training signal to alleviate the domain gap between video and sentence domains. In this paper, we propose a novel cross-domain alignment module to deeply align the features in two domains, yielding to learn better representations. We evaluate our model on three public datasets: Charades-STA, ActivityNet Captions and TACoS, and the experiments prove that our proposed cross-domain alignment module can bring significant improvement, and can be also applied to other baseline models.

Conventional Multi-Task Learning (MTL) models, such as hard sharing, adopt handcrafted network architecture, which shares entire layers for all tasks, and thus have two shortcomings: 1) negative transfer phenomenon and 2) low parameter efficiency. This paper proposes a novel neural network model, which allows different tasks to share a network at the parameter level. Specifically, the model defines a subnet for each task by adopting task-specific binary masks. The masks are trainable and can be learned together with network weights using standard back-propagation. Benefit from the fine-grained sharing mechanism, the negative transfer phenomenon can be alleviated, and the parameter efficiency is greatly improved. According to the experiments on a public dataset, our model outperforms the single-task baseline model even when only 0.8% of parameters remained in the subnets. Compared with the multi-task baseline model using fixed masks, our model is much more robust to changes in network sparsity.

XGNN: A chemometric dual-tower model for predicting aqueous solubility.

Intent recognition, as a core technology of human-computer interaction systems, directly affects the quality of intelligent services. Traditional single-modal intent recognition is limited by the constraints of a single information source, resulting in low accuracy in complex scenarios. This paper proposes a Cross-Modal Attention Fusion (CMAF) model for multimodal intent recognition, which achieves accurate intent recognition through deep fusion of text, audio, and visual multimodal information. The model adopts a hierarchical feature extraction architecture and combines intra-modal attention with cross-modal attention mechanisms to effectively capture semantic correlations and complementary information between modalities. Experimental results on three public datasets, namely MIntRec, SIMMC, and the newly added medical consultation dataset (MedDialog), show that the CMAF model improves the accuracy by 4.2%, 3.8%, and 5.1% respectively, and the macro-F1 score by 3.9%, 3.5%, and 4.7% respectively compared with baseline models. Particularly, the performance degradation of CMAF is only 5.2% under 0dB noise, verifying the significant advantages of multimodal fusion in intent recognition tasks.

Greenhouse gas (GHG) emissions from health care are substantial and disproportionately harm persons with cancer. Emissions from a central component of oncology care, outpatient clinician visits, are not well described, nor are the reductions in emissions and human harms that could be obtained through decentralizing this aspect of cancer care (ie, telemedicine and local clinician care when possible). To assess potential reductions in GHG emissions and downstream health harms associated with telemedicine and fully decentralized cancer care. This population-based cohort study and counterfactual analyses using life cycle assessment methods analyzed persons receiving cancer care at Dana-Farber Cancer Institute between May 2015 and December 2020 as well as persons diagnosed with cancer over the same period from the Cancer in North America (CiNA) public dataset. Data were analyzed from October 2023 to April 2024. The adjusted per-visit day difference in GHG emissions in kilograms of carbon dioxide (CO2) equivalents between 2 periods: an in-person care model period (May 2015 to February 2020; preperiod) and a telemedicine period (March to December 2020; postperiod), and the annual decrease in disability-adjusted life-years in a counterfactual model where care during the preperiod was maximally decentralized nationwide. Of 123 890 included patients, 73 988 (59.7%) were female, and the median (IQR) age at first diagnosis was 59 (48-68) years. Patients were seen over 1.6 million visit days. In mixed-effects log-linear regression, the mean absolute reduction in per-visit day CO2 equivalent emissions between the preperiod and postperiod was 36.4 kg (95% CI, 36.2-36.6), a reduction of 81.3% (95% CI, 80.8-81.7) compared with the baseline model. In a counterfactual decentralized care model of the preperiod, there was a relative emissions reduction of 33.1% (95% CI, 32.9-33.3). When demographically matched to 10.3 million persons in the CiNA dataset, decentralized care would have reduced national emissions by 75.3 million kg of CO2 equivalents annually; this corresponded to an estimated annual reduction of 15.0 to 47.7 disability-adjusted life-years. This cohort study found that using decentralization through telemedicine and local care may substantially reduce cancer care's GHG emissions; this corresponds to small reductions in human mortality.

Many relevant sound events occur in urban scenarios, and robust classification models are required to identify abnormal and relevant events correctly. These models need to identify such events within valuable time, being effective and prompt. It is also essential to determine for how much time these events prevail. This article presents an extensive analysis developed to identify the best-performing model to successfully classify a broad set of sound events occurring in urban scenarios. Analysis and modelling of Transformer models were performed using available public datasets with different sets of sound classes. The Transformer models' performance was compared to the one achieved by the baseline model and end-to-end convolutional models. Furthermore, the benefits of using pre-training from image and sound domains and data augmentation techniques were identified. Additionally, complementary methods that have been used to improve the models' performance and good practices to obtain robust sound classification models were investigated. After an extensive evaluation, it was found that the most promising results were obtained by employing a Transformer model using a novel Adam optimizer with weight decay and transfer learning from the audio domain by reusing the weights from AudioSet, which led to an accuracy score of 89.8% for the UrbanSound8K dataset, 95.8% for the ESC-50 dataset, and 99% for the ESC-10 dataset, respectively.

The existing knowledge graph-based recommendation models often lack a fine-grained consideration of collaborative information between users and items and overlook the high-order semantics and structural relationships within the graph paths. To address these issues, a dual-end recommendation algorithm integrating User Intent and Knowledge-Aware Graph Attention Networks (UIKGAN) is proposed. On the user end, the intent behind user-item interactions to refine the representation of collaborative information is modeled. By propagating relationship paths, UIKGAN aggregates deeper semantic and structural information from the knowledge graph to more accurately capture the extended representation of user intent and behavior patterns. On the item end, UIKGAN embeds and aggregates high-order neighboring triplet information using a knowledge-aware attention mechanism, enriching the feature representation of items. Additionally, this paper introduces an independence modeling module to optimize the loss function, providing better interpretability of user intent. Experiments were conducted on three public datasets, including comparative experiments with seven baseline models, ablation studies, hyperparameter sensitivity experiments, and sparse data issue analysis. The experimental results demonstrate that the UIKGAN model outperforms other baselines in overall performance, improving recommendation accuracy while effectively alleviating the issue of dataset sparsity.