
ABSTRACT Classroom monitoring presents substantial difficulties for object detectors deployed on teaching computers, including dense target distributions, pronounced within‐scene scale variation, and the high computational cost of conventional detection networks. This study develops LD‐YOLO, a lightweight detector based on YOLOv5s, through three targeted modifications. First, FasterBlock modules are introduced into the deeper P4 and P5 backbone stages to reduce redundant spatial computation, the parameter count, and computational complexity. Second, an ASFF‐3 module is incorporated after the neck to adaptively integrate the P3, P4, and P5 feature maps, thereby improving cross‐scale representation, particularly for small and scale‐varying targets. Finally, a scheduled ‐EIoU localization loss progressively adjusts the contribution of geometric errors during training to improve bounding‐box localization. On the private VOC_CBSD dataset, LD‐YOLO achieves an of 85.26%, which is 4.40 percentage points above the YOLOv5s baseline of 80.86%. On the public SCB‐Dataset3, its is 2.54 percentage points higher than that of YOLOv8n and 0.58 percentage points higher than that of YOLOv10‐n. Furthermore, CPU‐only evaluation on a teaching computer, using four inference threads and 640‐pixel inputs, yields an end‐to‐end throughput of 17.57 FPS. These results provide a reproducible assessment of the accuracy–efficiency trade‐off of LD‐YOLO on a public classroom‐behavior dataset and teaching‐computer hardware.
ABSTRACT Heart failure, a serious cardiovascular disease that demands precise and early evaluation of its severity for guiding efficient clinical decisions. Although few classification/risk assessment approaches are based on the image data but offer limited interpretation of echocardiography images. In this paper, a machine learning approach has been developed for automated estimation of heart failure severity from the Echography images of the EchoNet‐Dynamic dataset. The deep cardiac features are extracted using the pre‐trained EfficientNet‐B0, and feature selection is achieved using the Random Forest (RF), followed by optimization of the feature space with particle swarm optimization (PSO). The optimized feature vector is fed to the eXtreme Gradient Boosting (XGBoost) regression model to predict continuous severity values, and then mapped to the clinical interpretative severity‐stages (Mild, Moderate, Severe and Critical risk). The system additionally proposes temporal visualization and the NetworkX‐based visualization of stage transitions to provide better understanding of the predicted outcome. The achieved R score is 0.9995 with MAE of 0.1157, MSE of 0.0811, and RMSE of 0.2847, which showed the effectiveness of the proposed framework to effectively estimate the heart failure severity from the echocardiographic images and provide interpretative information of severity‐stage transition in an effective and comprehensible way. The developed framework can be a promising candidate for clinical decision support system. Because the EchoNet‐Dynamic dataset provides only cross‐sectional echocardiographic measurements without repeated patient follow‐up, the temporal ordering used for visualization is a simulated (pseudo‐temporal) construct designed to aid interpretability, rather than a model of observed longitudinal predicted severity‐state relationships.
ABSTRACT Online Social Networks (OSNs) are incredibly popular, acting as platforms where hundreds of millions of people share vast amounts of valuable data. Gender inference or prediction has been studied many times before due to its importance across various applications and is predominantly approached as a binary classification task. This study addresses gender inference in OSNs from an adversarial perspective, focusing exclusively on the use of usernames and display names. Unlike prior work that relies on extensive profile information or network features, the proposed approach operates under minimal and privacy‐relevant user data. Using data from Turkish Facebook users, two user subsets are extracted and used to construct five large‐scale datasets, each comprising 120K usernames and/or display names. A broad range of learning models, including traditional classifiers, deep learning methods, and transformer‐based architectures, is evaluated alongside multiple embedding techniques such as fastText. Experimental results demonstrate that a modified transformer‐based model, namely, Multi‐Pooling BERT classifier, achieves superior performance, attaining an F 1‐score of 0.952 on a hybrid dataset combining usernames and first names. This result represents a state‐of‐the‐art performance for machine learning–based gender inference of Turkish OSN users using only usernames and display names. To the best of our knowledge, this is the first study devoted to systematically investigating this problem setting, thereby providing both a novel contribution and a highly effective solution.
ABSTRACT Sophisticated code‐reuse and data‐oriented attacks continue to bypass traditional defenses, motivating practical and deployable protection mechanisms. This paper presents RACCO, an LLVM‐based compiler framework that provides a practical combination of stateful control‐flow integrity (CFI) and lightweight data‐flow integrity (DFI). RACCO combines a stateful CFI mechanism (R‐CFI) with transient call‐site‐sensitive validation to constrain indirect transfers, and a lightweight data‐flow integrity mechanism (L‐DFI) based on one‐bit taint tracking along explicit SSA dependencies to protect sensitive sinks. RACCO is implemented in LLVM 17 and evaluated against 18 real‐world CVEs and 225 synthetic attacks. Experimental results show mitigation rates of 83.3% and 97.3%, respectively, within the evaluated threat model, while evasions are limited to same‐relation redirections and uninstrumented boundaries. Across 120 benchmarks, RACCO introduces 5.3% geometric‐mean runtime overhead, with a low false‐positive rate, and preserves functional compatibility for supported user‐space applications. These results demonstrate that a practical, compiler‐based integration of CFI and DFI can deliver an effective and deployable defense against sophisticated exploit chains, without the prohibitive costs of more complex solutions.
ABSTRACT Age of information (AoI), which is defined as the time elapsed since the generation of the latest received update, has drawn growing research attention recently. To optimize the system AoI for computation‐intensive tasks, most existing work either follows simple queuing disciplines or works in an offline manner. Thus, in this paper, we investigate how to minimize the system AoI by taking into account both the freshness of the arrival packets and the computation load of future packets. First, we identify the existence of the Pareto‐optimal frontier for packet scheduling for the first time, which is obtained by defining two auxiliary attributes related to the freshness and the computation load of packets. Second, considering the possible future arrivals, we decompose the problem of AoI minimization into the subproblems of minimizing the possible incurred AoI at each epoch. Then, an efficient Pareto‐optimal packet scheduling method is proposed to minimize the incurred AoI with the selected packets. The proposed method is proven to have an approximation ratio of 5/3. Additionally, to further improve the performance, a Pareto‐optimal packet scheduling method with a retention mechanism is also proposed, which incorporates a candidate packet retention mechanism to dynamically maintain a queue of high‐quality candidate packets, considering the randomness of future packet arrivals. Finally, the theoretical analysis and simulation results verify that the proposed algorithm has high performance in terms of AoI, reducing the AoI by up to 12.6%.