Modern urban transportation and car-sharing systems, enabled by digital services, are being increasingly used. These services provide benefits for citizens and providers, but have to be carefully designed in order to be secure and privacy-enhancing. This work presents the analysis of privacy and security challenges in car-sharing platforms, specifically addressing issues, threats, and security challenges such as forensic readiness. We propose a novel and agile privacy-preserving solution tailored for car-sharing environments. Our solution integrates quantum-resistant primitives such as digital signature and group signature schemes, which preserves user privacy and protects against curious service providers and other third parties. Furthermore, the design incorporates forensic readiness by ensuring that despite privacy-preserving mechanisms, sufficient auditable data can be securely retrieved and analyzed in the event of security incidents or disputes, facilitating post-incident investigations without compromising legitimate user privacy. The presented experimental results of used cryptographic operations demonstrate practical applicability and efficiency, even for resource-constrained devices deployed in vehicles and users’ smartphones.

This study introduces a benchmark for evaluating information retrieval in the Slovak language, leveraging a question-answering dataset for fine-tuning and assessment of sentence transformers. The dataset, named Retrieval SkQuAD, is integrated into two widely recognized evaluation frameworks: BEIR (Benchmarking Information Retrieval) and MTEB (Massive Text Embedding Benchmark). Retrieval SkQuAD (Slovak Question Answering Dataset) comprises 19,000 manually annotated answers to 1,134 questions, with each answer assigned a relevance score, and includes information on the usefulness of documents in generating responses. Unlike question answering datasets, this resource provides a nuanced assessment of partial relevance across multiple documents. We fine-tuned several sentence transformers and BERT-based models specifically for retrieving documents containing correct answers within the Slovak Wikipedia. Our fine-tuning process incorporates adversarial questions as hard negatives, leading to significant improvements in retrieval accuracy. Experimental results demonstrate that our approach advances state-of-the-art performance for Slovak information retrieval.

Missing data in clinical time series is pervasive and decision-critical, arising from irregular sampling, workflow-driven measurement policies, sensor failures, and intervention-dependent monitoring. Despite extensive methodological work on imputation, practitioners still lack clear guidance on selecting approaches that are safe, interpretable, and feasible in real clinical systems. This framework presents a decision-centric view of clinical time series imputation that links missingness mechanisms--- i.e ., missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR)---as well as temporal irregularity, nonstationarity, noise, class imbalance, and governance constraints to defensible method selection. We show that low reconstruction error does not guarantee clinical safety: imputation can attenuate rare events, distort uncertainty, and misalign with decision thresholds even when root mean squared error (RMSE) or mean absolute error (MAE) improves. Rather than proposing a new algorithm, this work provides practical guidance on when imputation is identifiable, when uncertainty or abstention is preferable to point reconstruction, and how evaluation must extend beyond point-wise accuracy to downstream decision impact. The framework aims to support safer integration of imputation into real-world clinical monitoring and decision support systems.

Hard-label black-box adversarial attacks on text classification models present significant challenges due to both the discrete, non-differentiable nature of text data and the lack of direct access to model predictions. Existing methods remain limited in efficiency and effectiveness—often becoming trapped in local optima or requiring excessive queries—which hinders the generation of high-quality adversarial examples with preserved semantic coherence under query constraints. To address these limitations, this study proposes KEM-Attack: a knowledge-enhanced metaheuristic framework that efficiently generates adversarial examples through synergistic integration of linguistic knowledge and metaheuristic optimization. KEM-Attack pioneers a two-stage methodology: first, constructing a knowledge-enhanced substitution space using WordNet’s taxonomic relationships and HowNet’s fine-grained sememe representations; second, introducing the first adaptation of Grey Wolf Optimization (GWO) to textual adversarial attacks, which is augmented with hierarchical position updating and adaptive mutation strategies. Extensive experiments show that KEM-Attack outperforms advanced baseline methods in terms of attack success rates, semantic similarity, and perturbation rates, while also improving query efficiency. The proposed approach demonstrates how integrating linguistic knowledge with metaheuristic optimization can significantly enhance adversarial attack performance, providing a deeper understanding of neural model vulnerabilities in natural language processing (NLP).

Breast cancer remains one of the principal reasons of death among women worldwide, highlighting the critical requirement for primary and exact diagnosis to expand patient persistence rates. Manual clarification of histopathological slides is frequently subjective, time-consuming, and prone to inter-observer unpredictability. Recent developments in machine learning (ML) and deep learning (DL) techniques have renovated histopathological image analysis by qualifying automated, consistent, and reproducible diagnostic systems. This review analysis thoroughly observes research from 2018 to 2025, covering advanced methods for breast cancer detection based on histopathological images, with feature extraction, feature selection, and classification strategies, and medical image investigations. In this study compares existing ML methods, convolutional neural networks (CNN), vision transformers (ViT), graph convolutional networks (GCN), ensemble learning and hybrid frameworks across numerous standard and multiple benchmark datasets. Performance comparisons encompass filter-based, wrapper-based, hybrid, and ensemble learning models. In this study highlights the benefits of AI-driven diagnostic systems such as enhanced accuracy, scalability, and consistency, while addressing key challenges including data discrepancy, inadequate interpretability, and high computational cost. This comparative review highlights publicly accessible datasets, standard architectures, and emerging techniques in explainable artificial intelligence, ViT, and federated learning. Finally, potential directions for future research are drawn to guide researchers, clinicians, and developers in the direction of proceeding intelligent systems for breast cancer histopathological diagnosis.

Apple leaf diseases (ALD) pose a significant challenge to global apple production, and accurately identifying ALD is crucial for reducing pesticide use and improving fruit quality, particularly in the context of smart agriculture. However, traditional approaches rely on single-model feature extraction, failing to account for relationships between different tasks, which limits their applicability in the apple industry. To address this, we design an optimized convolutional neural network–vision transformer (CNN–ViT) hybrid framework named ResViT-AM, focusing on domain-specific enhancement rather than architectural novelty. Instead of proposing a completely new structure, this work refines existing CNN–Transformer paradigms through task-oriented feature fusion and adaptive attention weighting, tailored for apple leaf disease classification under complex orchard conditions. Using a weighted attention fusion mechanism, our model dynamically integrates features extracted by Residual Network 101 (ResNet-101) and vision transformer (ViT), combining proven architectures in a task-adaptive way rather than pursuing architectural innovation, blending the local convolutional details of Residual Network (ResNet) with the global contextual features of ViT. This approach enhances the model’s representation capability and allows parallel processing of multiple tasks, thereby saving training time and computational resources. Additionally, we evaluate on the public AppleLeaf dataset, which reflects real-world outdoor conditions. On its held-out test split, our model achieves 99.14% top-1 accuracy on the AppleLeaf test split, indicating promising performance under complex orchard conditions. Compared with representative convolutional baselines, ResViT-AM shows greater stability and adaptability on challenging cases, offering a competitive and practical solution for automated apple leaf disease diagnosis.

Heterogeneous application environments necessitate the coexistence of diverse Congestion Control Algorithms (CCAs). However, traditional active queue management (AQM) schemes are typically agnostic to specific CCAs and treat heterogeneous traffic flows uniformly, thereby overlooking the diverse buffering requirements of these CCAs. Consequently, bandwidth allocation does not explicitly consider latency and buffer utilization metrics. We present Niche, a lightweight, contention-aware AQM designed for programmable switches. Niche leverages local buffer—backlog statistics to dynamically classify and isolate flows into physical queues without incurring complex measurement overhead. Combined with dynamic bandwidth allocation, Niche not only improves inter-flow fairness but also optimizes comprehensive performance. Experimental validation on programmable switches confirms that in multi-CCA coexistence scenarios, Niche reduces total buffer occupancy by 50% while stabilizing the Jains’ Fairness Index at 0.98. Crucially, it lowers end-to-end latency for latency-sensitive flows like BBR and Vegas by an order of magnitude. Niche approximates the fairness of Fair Queuing (FQ) with limited queue resources, outperforming comparable baselines and demonstrating consistent performance gains as queue availability scales.

With the rapid development of artificial intelligence generation technology, the boundary between artificial intelligence (AI)-generated images and real images is becoming increasingly blurred, posing serious challenges to the credibility and authenticity of digital content. Addressing the insufficient generalization of existing AI-generated image detection methods in complex scenarios, this research proposes a Diffusion-Cross Attention Transformer (DCAT) framework for image authenticity verification. This framework innovatively combines diffusion model feature extractors and cross-attention vision transformers (ViT) to achieve fine-grained capture of image microscopic noise distribution and semantic relationships. Large-scale experimental validation was conducted on the GenImage dataset. The model demonstrated excellent performance in various degradation environments, with area under the receiver operating characteristic curve (AUC) remaining stable from 0.910 under no degradation conditions to 0.775 in extreme degradation environments, significantly outperforming traditional methods. The core contributions of this research include proposing a multi-scale noise analysis feature extraction method, constructing a cross-attention semantic association detection mechanism, and theoretically deepening the mathematical characterization of distribution differences between generated models and real images. This innovative approach not only provides key technological breakthroughs but also offers important technical support for maintaining the authenticity of digital content ecosystems, holding significant scientific and practical value for the field of artificial intelligence image generation and detection.

This article introduces MinMAE, a novel activation calibration method for Post-Training Quantization (PTQ) that significantly reduces accuracy loss in Convolutional Neural Networks (CNN). Motivated by the need for high-fidelity quantization without costly retraining, MinMAE directly minimizes the Mean Absolute Error (MAE) between original and dequantized activations, making it robust to outliers that degrade standard methods. We evaluated MinMAE against absmax and mean+3std on YOLOv8n and YOLOv8s models. For 8-bit quantization on YOLOv8n, MinMAE achieved 8.4% mean Average Precision (mAP) loss, whereas standard methods led to drops of up to 14.4%. On the larger YOLOv8s model, MinMAE also demonstrated superior performance, reducing loss from 7.2% to just 5.2%. While this approach requires a higher one-time calibration time, the substantial accuracy gains establish MinMAE as a highly effective and broadly applicable strategy for improving the stability and performance of quantized CNNs.

Electroencephalography (EEG) signals are physiological signals that directly reflect the brain’s electrical activity, holding significant research value in the field of emotion recognition. However, existing methods still face certain limitations in jointly modeling temporal, spatial, and frequency-domain features. To address this, this article proposes an EEG-based emotion recognition model that integrates a three-dimensional convolutional neural network (3DCNN) with a bidirectional long short-term memory network (BiLSTM), referred to as 3DC-BiL. This approach utilizes 3DCNN to extract spatio-temporal features across different frequency bands, and then leverages BiLSTM to enhance temporal modeling capabilities, achieving collaborative fusion of multidimensional features. On the DEAP dataset for binary emotion classification, the proposed model achieves average accuracies of 98.60% and 98.66% on the arousal and valence dimensions, respectively. On the SEED dataset for three-class emotion classification, it achieves an average accuracy of 99.01%, outperforming various mainstream emotion recognition models in recent years. The experimental results demonstrate that the proposed model excels in EEG feature extraction and high-precision emotion classification, providing a new perspective and practical foundation for the research of EEG-based emotion recognition systems.