Adverse event detection for non-pharmaceutical substances, such as dietary supplements, herbal remedies, and alternative therapies, remains a major blind spot in current pharmacovigilance systems, which are primarily designed for prescription drugs. This research addresses that gap by introducing a high-precision matching framework that links scientific literature from PubMed with real-world safety reports from the FDA Adverse Event Reporting System (FAERS). The proposed approach incorporates a robust pre-processing pipeline, followed by a threelayer matching engine based on rule-based substance overlap, metadata alignment, and temporal validation. The result is a curated set of 560 article-report pairs obtained by applying our multi-stage matching pipeline to two datasets: a filtered subset of 38,978 PubMed articles that mention non-pharmaceutical substances, drawn from the 2024 PubMed baseline distribution and 135,848 adverse event reports from the FAERS for Q3 2024 (JulySeptember). This represents the first structured dataset connecting peer-reviewed research with reported adverse events for nonpharmaceutical compounds. Experimental results show that while raw rule-based matching produced 670 links, our enhanced pipeline reduced noise and improved semantic accuracy, prioritizing reliability over quantity. This dataset lays the groundwork for safer, data-driven integration of alternative medicine into modern pharmacovigilance efforts. Moreover, results so far serve as an important first step toward making available a benchmark dataset for evaluating non-pharmaceutical adverse event extraction and article-report linking systems. This work provides a reproducible, high-precision framework for linking literature and adverse event reports, addressing challenges not covered in prior studies.
The absence of debugging symbols in stripped binaries, particularly descriptive function names, hinders reverse engineering and security analysis. This paper introduces a structured pipeline for automated function name inference, combining Transformer models, calibrated confidence estimation, and context-aware refinement using large language models pretrained on code. In the first stage, a specialized Transformer model is trained on normalized assembly instructions to predict initial function names, along with statistical indicators. These indicators are then mapped to reliable confidence scores using a logistic regression calibration module. In the final stage, lowconfidence functions are reprocessed using fine-tuned, pretrained code language models, which leverage the context derived from high-confidence caller and callee relationships. The method is validated on a dataset of approximately 21.3 million functions extracted from 6,824 Windows executables compiled with various MSVC versions and architectures. The results demonstrate that our approach supports the practical recovery of meaningful function identifiers and enhances malware analysis tasks.
In this work, we propose an innovative knowledge distillation framework tailored for promptable transformer-based segmentation models, focusing on Segment Anything Model 2 (SAM 2). Powerful models, such as transformer-based ones, which require large-scale datasets and extensive computation for good performance, are difficult to train, even with limited data, often yielding uncertain segmentation results across different use cases. We address this challenge by proposing a distillation strategy that allows a student model to inherit structured knowledge from a strong pre-trained teacher with the same architecture, even when trained on small datasets. The method combines hard supervision from ground-truth masks with soft guidance from the teacher’s logits, helping the student learn meaningful structures early in the training process. To address the challenge of noisy or overconfident predictions from the teacher, we introduce a confidence-aware agreement mask that modulates the soft loss and provides guidance for the student to focus on regions where the teacher is more confident. The work is extended with token-level distillation to enhance structural fidelity in student representations. Experiments demonstrate the effectiveness of our method in low-data regimes, with the student model achieving a notable reduction in the total loss (from 15.17 to 8.84) and up to $20 \%$ improvement in Intersection-over-Union (IoU) across various inference scenarios. Qualitative results further support these findings, showing visibly cleaner segmentation outputs with reduced noise, indicating improved generalization under data-scarce conditions.
Fine-grained change detection in structured remote sensing environments, such as agricultural fields, requires models that can reason over semantically coherent regions while preserving spatial precision. Existing methods often rely on convolutional or transformer-based architectures with uniform tokenization, limiting their ability to capture changes aligned with realworld semantic structures. In this paper, we propose SFCNet, a region-aware change detection framework that leverages explicit field-level segmentation masks to guide both token aggregation and spatial decoding. Our method introduces a Field-Aware Region Encoder (FARE) that aggregates features within each semantic region, producing interpretable region tokens. These are processed through a slot attention mechanism to model temporal differences across structured units. A re-projection decoder then restores field-wise predictions to pixel-space using the original geometry. We evaluate SFCNet on the CLCD dataset, where it outperforms prior state-of-the-art models by $\mathbf{+ 3. 3 1 \%}$.
Online platforms are complex systems that influence the commercial, social, and political environment, debating important real-life topics, e.g., health, emigration, elections, climate change, etc. These online environments offer users freedom of expression through anonymous posting. In addition to their obvious advantages, some users abuse this freedom to spread harmful content, e.g., misinformation, propaganda, harmful conspiracy theories, or abusive, aggressive, and offensive speech. Automated detection techniques can effectively reduce the negative influence of antisocial behavior used by these malicious actors. In this article, we propose HarmLLAMA, a fine-tuned LLAMA2 model using LORA. The experimental results on two real-world datasets show that our model, HarmLLaMA, outperforms current state-of-the-art models in terms of Accuracy, Precision, Recall, and F1-Score.
Automated multi-label electrocardiogram (ECG) diagnosis promises accessible cardiovascular care, but clinical adoption is challenged by deep learning models’ size, interpretability, and poor performance on rare conditions. This paper presents an end-to-end framework addressing these issues through two novel, compact convolutional neural network (CNN) architectures—an Efficient CNN ($\sim 20 \mathrm{k}$ parameters) and a Compact MiniResNet $(\sim 35 \mathrm{k}$ parameters)-trained on the 21,837-record PTB-XL dataset. Our methodology integrates (1) a robust class weight strategy for label imbalance, (2) waveform-aligned Grad-CAM for clinically relevant visual explanations, and (3) post-training, per-class decision threshold optimization to significantly improve sensitivity on minority classes. Models are validated for edge deployment via full INT8 quantization. Our compact models achieve diagnostic performance comparable to multi-million parameter ResNet baselines (MiniResNet Macro-AUC of 0.961, Macro F1-score of 0.70). Threshold optimization boosted recall on the rarest class (Hypertrophy) from 0.17 to 0.52. Post-quantization, model size is reduced by $87 \%$ (to 65 KB) with negligible performance drop, and inference latency by $18 x$ on a CPU, confirming edge suitability. This work provides a reproducible blueprint for developing trustworthy, lightweight, and interpretable AI-powered cardiology assistants.
This paper investigates the benefits of further fine-tuning the HateBERT model for hate speech classification, addressing the challenges posed by class imbalance and linguistic subtleties inherent in hateful content. The fine-tuning process employs a gradual unfreezing schedule combined with a custom Focal Loss function to effectively handle imbalanced datasets and mitigate overfitting. Training is optimized using the AdamW optimizer with linear learning rate scheduling and gradient accumulation to maintain stable convergence. Evaluation on representative hate speech examples shows that the fine-tuned HateBERT achieves higher classification confidence compared to the original pre-trained model, improving the detection of subtle and implicit hateful expressions. Additionally, an LLM (Large Language Model)-based post-evaluation using Mistral-7B provides interpretable justifications and expert-style validation of the model’s predictions, enhancing transparency and trustworthiness. These results underscore the importance of task-specific adaptation and multi-model interpretability approaches to improve performance on complex and nuanced NLP tasks such as hate speech detection.
Perception in aerial environments depends on object detection and semantic segmentation to achieve scene understanding. However, the limited availability of semantically annotated datasets poses a challenge due to the extensive manual effort needed for their creation. This work proposes a customized pipeline designed to generate high-quality semantic masks from object detection bounding boxes by integrating the Segment Anything Model (SAM), a suite of image processing techniques, and adaptive prompting methodologies. The input prompts represent either foreground objects or background regions and are refined through additional image processing operations to improve quality and mask accuracy. The primary contribution of this work is the use of point-based prompts to accurately guide the class-agnostic SAM model in delineating precise semantic boundaries. Furthermore, a genetic algorithm is employed to intelligently select a reliable and informative set of object points, thereby enhancing the robustness and accuracy of the segmentation process. The proposed method was evaluated using a manually annotated subset of the AU-AIR dataset, focused on vehicle instances, achieving a segmentation score of 84.62% IoU. This represents an 18% improvement over the baseline UNetFormer model trained on the UAVid dataset.
In this paper, we propose a solution that leverages knowledge graphs to model and compare cooking recipes, aiming to bridge the gap between free-form user instructions and a structured recipe database. By integrating a hierarchical, domain-specific knowledge graph with inductive graph convolutional networks (for scalable embedding) and a fine-tuned multilingual transformer (for precise information extraction), our approach excels in both efficiency and matching accuracy. We achieve 97.91% accuracy on embedded recipes within the semantic network, yielding a maximum similarity of 0.54 for different versions of the same dish, and reach up to 0.32 similarity for dishes not present in the database. The motivation for this work lies in the need for end users, especially in low-resource or multilingual settings, to retrieve reliable, relevant recipes without manual browsing or query formulation. Our system addresses this by parsing user-provided recipe text in six languages, embedding both user and candidate recipes into a shared vector space, and computing similarity results to return the best match.
Electroencephalography (EEG) remains one of the most widely used non-invasive methods for investigating brain function, due to its high temporal richness and accessibility. Still, the complex and noisy nature of EEG signals makes decoding neural states a challenging task, particularly when using traditional signal processing or shallow machine learning approaches. In recent years, deep learning models, such as Transformers, have emerged as a powerful alternative to classical machine learning methods, thanks to their capacity for modeling long-range dependencies and learning contextual representations. In this paper, we explore the application of Transformer architectures to EEG data for the purpose of binary and three-class classification in a visual perception task. To evaluate model performance, we compare the Transformer against four other models: XGBoost, Random Forest, a simple Multi-Layer Perceptron and a Long Short-Term Memory network. Our findings suggest that attention-based models are capable of capturing informative temporal dependencies in EEG signals, with preprocessing playing a critical role in optimizing representation quality. This work highlights the potential of Transformer models for non-invasive neural decoding and contributes to the development of more effective tools for cognitive neuroscience and brain-computer interfaces.
This paper presents efforts to enhance driving safety through the analysis of on-road behavior and personalized driver feedback. It reviews recent literature focused on smartphonebased driving assistance applications, highlighting trends in sensor-driven behavior monitoring and in-app feedback. The hybrid mobile application DriverTracker, developed using the Ionic framework, is analyzed, detailing its architecture, implementation, and driver scoring algorithm. Experimental results suggest that DriverTracker can identify unsafe driving behaviors by integrating sensor data with contextual user input before and after the trip. While the current study is limited in scope, it offers a proof of concept for combining subjective and objective metrics in driver behavior analysis.
Automatic detection of epileptic seizures based on EEG signals represents a major challenge in neuroscience and AI-assisted medical applications. Currently, over 50 million people worldwide live with epilepsy, and a significant proportion of them experience recurrent seizures that can occur unpredictably and severely affect their quality of life. For this reason, we propose EnfXAI, a hybrid system that integrates a deep learning model (CNN), an explainability component (SHAP), and a fuzzy logic module for risk estimation. The CNN model is trained on an EEG dataset segmented into one-second windows, achieving an accuracy of over 98% in classifying seizure and non-seizure windows. SHAP is then used to identify the EEG features with the highest positive contribution to seizure prediction. These features are subsequently processed by a fuzzy logic system that provides an interpretable seizure risk estimate—classified as low, medium or high. In this way, we contribute to the early and interpretable identification of epileptic seizures, helping affected individuals benefit from faster and more personalized interventions. As a result of the analysis, the test data was split into two categories: at-risk (693 individuals) and not at-risk (494 low-risk and 1113 medium-risk individuals). The final output is both a highly accurate and easy-to-understand prediction, with strong potential for application in real-world clinical contexts. The proposed system represents a significant step toward integrating explainable AI (XAI) into neurological diagnostics.
This paper introduces a novel approach for speaker anonymization based on cascaded speech-to-text (STT) and text-to-speech (TTS) Deep Neural Networks (DNNs), where the multi-speaker TTS component is driven by a Generative Adversarial Network (GAN) to anonymize the speakers’ identity and by a prosody transfer module to preserve the original prosody. The GAN creates artificial speaker embeddings that do not correspond to any real identity, enabling a high diversity of synthesized voices. The prosody preservation and speaker embedding generation for the TTS component are clearly distinct approaches: the embeddings are progressively generated from a low and compact latent space, while the prosody is directly cloned onto the synthesized speech. The experimental results show low cosine similarity between original and anonymized embeddings, with some minor degradation in linguistic fidelity due to the performance of the STT and TTS components. Additionally, the developed visual interface supports comparative analysis of audio features before and after anonymization, contributing to a qualitative assessment of system performance.
Low-resource languages face significant challenges in the development of robust natural language processing (NLP) tools due to the scarcity of standardized corpora and linguistic resources. Tokenization is a critical step in natural language processing, yet existing tokenization algorithms often struggle with agglutinative and low-resourced languages due to their rich morphological structure and limited training data. In this paper, we present the creation of a corpus for Crimean Tatar, a severely under-resourced Turkic language. To address orthographic inconsistencies and dialectal spelling variations, we utilize bilingual dictionaries to normalize and consolidate different variants of the same word. Leveraging the linguistic similarity between Crimean Tatar and Turkish, we utilize a large Turkish corpus alongside a smaller Crimean Tatar corpus to adapt and improve subword tokenization algorithms such as Byte-Pair Encoding (BPE), WordPiece, and Unigram. We present a fine-tuning algorithm that incorporates a golden set of annotated tokenizations and a curated Crimean Tatar vocabulary to adjust token boundaries. To further address context-sensitive morpheme boundaries, where some morphemes should be merged or separated depending on usage, we model tokenization as a character-level sequence labeling task. A BiLSTM model trained on the golden dataset is introduced, assigning “$\mathbf{B}$” (begin) and “$\mathbf{I}$” (inside) labels to characters to predict token splits. This approach demonstrates improved handling of context-dependent segmentation and outperforms traditional tokenization methods. We provide detailed analyses of the corpus composition, normalization strategies and tokenizers creation and evaluation. Our results demonstrate that a larger corpus of a very similar agglutinative language, in this case Turkish, can improve the performance of the tokenizer, but more important is the usage of a vocabulary. However, a model which takes into account the context of characters can better tokenize words composed of morphemes than tokenizers based on learning merging rules.
This paper presents the design and implementation of a stealthy dynamic analysis environment for Linux, named Rootalyx. The system addresses the challenge of evasive malware that can detect and circumvent traditional monitoring tools. Its core component is a kernel module that uses the ftrace subsystem to instrument system calls, integrated with rootkitinspired hiding techniques to ensure a covert operational profile. The collected behavioral data, including system call arguments, is processed by a hierarchical hybrid model combining the LSTM and Transformer architectures for automated malware classification. Experimental results show that the proposed system introduces lower performance overhead compared to traditional tools for monitoring like strace and successfully evades detection by common security scanners. The machine learning model achieves a classification accuracy of 80.34% on the extracted data, validating the viability of kernel-level stealth monitoring for robust malware analysis.
Image security plays a crucial role in various domains where sensitive information is transmitted, such as telemedicine, satellite imagery transfer, and military communications. Effective image encryption requires strong security, secure network transmission, and low computational cost. Traditional encryption methods provide robust security and reliable transmission but are computationally expensive. Chaos-based cryptosystems offer lower computational costs but generally lack sufficient network security. Experimental neural networkbased encryption achieves high processing speed but introduces overhead due to model training and is limited by its dependence on specific datasets, making it non-adaptive.In this work, we propose a novel image encryption scheme combining chaotic theory, elliptic curves, and static S-boxes. The method initially employs the CCS chaotic map to generate a master key sequence used for encrypting the plain image. These static keys are further utilized in a symmetric process involving secret key derivation based on scalar operations over elliptic curves and S-box transformations. Experimental analysis demonstrates that the proposed scheme offers strong resistance to side-channel and statistical attacks, including differential analysis, outperforming several existing methods in terms of security. The main limitation lies in the increased computational cost due to elliptic curve operations used in the key derivation process.
This paper presents an integrated medical platform that streamlines healthcare workflows by eliminating reliance on traditional document-based methods for handling patient records. Building on previous work on structuring unstructured medical data, the proposed extension now enables direct generation, editing, and storage of Electronic Health Records (EHRs) within a secure web interface. The platform supports dynamic form creation, customizable templates, rolebased access, and cloud-based synchronization. To enhance usability, an intelligent assistant is integrated using a retrievalaugmented generation (RAG) pipeline. A locally trained NLP module handles intent classification and slot extraction, retrieving structured data from the medical database. Outputs are refined using OpenAI to deliver coherent, context-aware responses. The platform demonstrates improved efficiency, data consistency, and user interaction, aligning with digital health goals and EMRAM Stage 6+ criteria.
Wearable sensors (e.g., accelerometers, gyroscopes) combined with machine learning (ML) methods offer a promising solution for remote monitoring of physical rehabilitation exercises. However, challenges arise given the high variability in how patients perform the same exercises, combined with the scarcity of labeled data. This limits the effectiveness of standard supervised learning approaches. This paper analyzes and quantifies key aspects of developing generalizable physical exercise recognition models through the lense of ML training procedures. We make four contributions: (i) evaluating time-domain data augmentations on model performance, (ii) quantifying inter-person variability using within-dataset incremental fine-tuning, (iii) assessing transferability across datasets with encoders pretrained in a supervised manner, and (iv) comparing supervised and unsupervised pretraining strategies using free-living and exercise-specific motion data.
When implementing an automated support system, organizations tend to rely on complex, computationally intensive architectures. This paper presents the results we obtained when exploring the possibility of developing a cost-effective question answering framework for this purpose. We explored how various models, context configurations, and personas (instructions that guide the system in terms of answer purpose and/or tone) impact customer satisfaction regarding the respective answers. Three Llama-3, 7 billion parameter version variants, fine-tuned across different combinations of training parameters, extra context provision and personas were evaluated. The solution obtained was implemented as a chatbot focused on questions regarding a specific university. Its performance was assessed based on the customer satisfaction scores obtained from university students. Results demonstrate that fine-tuned models with 1000 iterations (the point at which the model starts to overfit slightly) achieved up to 85% customer satisfaction when combined with both extra context and persona features, compared to 25% for the base model alone. These findings suggest that effective specialized chatbots could be implemented without complex Retrieval Augmented Generation architectures and provide practical guidance for organizations which aim for simplicity, cost-effectiveness and replicability in their systems’ development.
Security operations centers (SOCs) are increasingly deploying large language models (LLMs) to assist human analysts; however, the reliability of off-the-shelf LLMs for malware analysis remains uncertain. This paper investigates whether state-of-the-art LLMs can serve as “virtual malware analysts” without task-specific fine-tuning. We introduce ArA (Artificial Analyst), a novel system that combines dynamic malware sandboxing with an LLM-based reasoning engine. ArA executes programs in a controlled environment, translates the observed behaviour into natural-language reports, and prompts an LLM to classify the behaviour as malicious or benign. We evaluated 12 modern LLMs-including GPT-4, its distilled variants, open-source models (1B-27B parameters), and a security-tuned model-on a dataset comprising 712 real malware samples and 304 benign programs. GPT-4 achieved the highest accuracy at 95% (macro F1 = 0.93), correctly identifying nearly all malware samples and rarely mislabeling benign software. Smaller open-source models demonstrated mixed performance; some attained perfect malware recall but exhibited higher false-positive rates on benign data, while the security-tuned 8B model performed below general-purpose models. We analyse each model’s strengths, common errors (e.g., false-positive biases), and the usage of tentative (“suspicious”) versus definitive labels. Our findings demonstrate that advanced LLMs can accurately distinguish malware behaviours in zeroshot settings, offering immediate value for SOC triage. However, we also highlight limitations, including insufficient coverage of behaviour variations and limited explanation quality, and discuss future steps to enhance LLM-based malware analysis for practical deployment.
