Many non-verbal individuals depend on symbol-based communication, yet AAC devices are frequently abandoned because they are bulky, conspicuous, or hard to reconfigure. Cue Clear addresses this with a wearable display that cycles through caregiver-curated communication symbols, controlled by a companion smartphone app. The system pairs a CircuitPython microcontroller driving an ILI9341 LCD—hosting a non-blocking HTTP server over a Wi-Fi access point—with a cross-platform Flutter application that builds symbol “profiles” and activates them on the device [3]. Key challenges included reliable communication with a resource-constrained device, responsive image browsing, and an accessible caregiver interface; these were met with a non-blocking embedded server, a read-through image cache, and a largetarget, confirmation-guarded UI. In testing, caching cut image load latency about 27-fold to 43 ms, and index-only profile activation stayed near 215 ms regardless of profile size. By directly targeting known causes of AAC abandonment, Cue Clear offers a sustainable, low-friction communication aid.
Social-media steganography is a hostile-channel problem because platforms such as Facebook can resize images, recompress them as JPEG, modify chroma information, remove metadata, quantize colors, and apply proprietary optimizations that destroy fragile hidden payloads. This paper presents FBStegNet, an end-to-end deep neural framework for robust binary data hiding in color images transmitted through Facebook-like processing. The key contributions are fourfold: a dense message encoder that distributes payload redundancy across the image, a residual CNN feature extractor that identifies stable cover structures, a differentiable Facebook simulation layer that models randomized JPEG compression, downsampling–upsampling, color quantization, and signal-dependent noise, and a texture-aware attention embedding network that controls where and how strongly information is embedded. By training the en coder and decoder over a stochastic distribution of social-media distortions, FBSteg Net optimizes message recovery after platform processing rather than only preserving clean-image fidelity. In a representative 0.05 bpp case study, FBStegNet reduces the bit-error rate (BER) to 0.018 in simulated Facebook-like tests, compared with 0.482 for LSB matching, 0.214 for DCT robust embedding, and 0.167 for generic neural hiding, while preserving 38.94 dB PSNR and 0.958 SSIM. In real Facebook upload download experiments, it achieves 0.027 BER with 38.51 dB PSNR and 0.951 SSIM, demonstrating that the learned embedding remains recoverable under practical social media transformations. Additionally, this work specifies a reproducible evaluation protocol using steganography image benchmarks, Facebook-style simulated channels, real upload–download tests, payload metrics, perceptual-quality measures, and steganaly sis checks. FBStegNet therefore provides a practical and extensible design for robust secret message communication over lossy social-media image pipelines. Keywords: Information hiding, social media, deep learning, image steganography.
Accurate drone detection in aerial and ground-based surveillance remains difficult because small unmanned aerial vehicles often appear as low-resolution, low-contrast targets embedded in cluttered backgrounds. Their visual appearance changes rapidly with altitude, viewpoint, illumination, motion blur, compression, and weather, causing many general-purpose detectors to lose fine spatial evidence or produce false alarms from birds, towers, clouds, and building structures. This paper proposesed a Multi-Scale Deep Learning Framework (MSDLF) for robust drone detection and classification in complex aerial scenes. The framework integrates a residual convolutional backbone, feature-pyramid aggregation, adaptive scale-attention fusion, decoupled detection heads, and multi scale consistency regularization. Unlike fixed pyramid fusion, the proposed attention module learns image-conditioned weights for features extracted at different strides, allowing the detector to emphasize high-resolution maps for distant drones while preserving deeper semantic context for clutter rejection. Experimental evaluations on VisDrone aerial images and UAV video benchmark datasets demonstrate that MSDLF achieves superior detection performance compared with two-stage, one stage representative, anchorfree, and transformer-based object detection methods. The experimental results of comparison, MSDLF achieves 92.1% precision, 90.4% recall, 91.2% F1-score, 94.7% AP50, and 71.9% mAP, improving over YOLOv8 by 3.4 percentage points in precision, 4.2 points in recall, and 4.1 points in mAP. Ablation and robustness studies further show that multi-scale fusion, adaptive attention, and consistency learning each contribute to higher small-object recall and better stability under blur, low light, compression, and scale reduction. These results indicate that scale-aware attention is an effective and practical mechanism for drone detection systems that must balance accuracy, robustness, and near-real-time inference.
Security Operations Centers rely on detection engineering to convert telemetry from identity systems, endpoints, networks, and applications into actionable security alerts. However, many organizations struggle to align detections with adversary behavior, measure coverage, and maintain detection content at scale. This paper presents a practical ATT&CK-aligned detection engineering framework that combines telemetry assessment, field analysis, detection design, validation, tuning, and continuous monitoring. The framework is demonstrated through case studies covering password spraying, excessive DNS queries and failures, suspicious PowerShell activity, and Detection-as-Code workflows. Each case study maps detection logic to relevant MITRE ATT&CK techniques and discusses data sources, validation methods, tuning considerations, and operational lessons. The paper also compares the proposed approach with traditional indicator-based detection, standalone SIEM rule development, and ATT&CK mapping used only for reporting. The results show that ATT&CK-aligned detection engineering can improve visibility, reduce unmanaged detection gaps, and support scalable security monitoring across healthcare, financial services, e-commerce, education, government, and artificial intelligence environments.
The rapid growth of visual databases calls for efficient Content-Based Image Retrieval (CBIR). Texture descriptors are central to these systems; however, their performance often degrades under geometric image transformations, particularly rotation. This paper presents a CBIR framework designed to compare handcrafted and deep texture features for rotation-invariant retrieval. A novel hybrid approach combines Local Binary Patterns (LBP) with the Stationary Wavelet Transform (SWT) to extract compact, multi-scale descriptors robust to orientation variability. In parallel, a transfer learning strategy leverages intermediate layers of pretrained convolutional neural networks (VGG16 and ResNet50) with multi-angle feature aggregation to extract rotation-robust deep descriptors. Experiments on benchmark texture datasets (Outex and Kylberg) show that the deep transfer-learning approach achieves higher recall at the cost of larger descriptor dimensionality and greater computational and memory demands, whereas the proposed hybrid descriptor provides a favorable trade-off between accuracy, compactness, and computational efficiency, making it well-suited for resourceconstrained applications.
Mountain biking carries significant injury risk, with up to 90% of cross-country riders sustaining at least one injury per season. A major cause is improper suspension setup, which is technically complex and difficult for everyday riders to optimize. This paper develops a crossplatform mobile application that combines a Raspberry Pi sensor unit with an AI advisor to deliver personalized suspension recommendations. The app, built in Flutter and Dart, streams live suspension data from the bike over Bluetooth Low Energy, sends rider profiles to OpenAI's GPT-4o-mini model via a REST API, and displays personalized trail recommendations using Google's Geocoding API. We tested the AI's accuracy by comparing 10 rider profiles against the official Fox 36 and RockShox setup charts, finding the AI produced recommendations within Fox's ±10 PSI tolerance on 70% of profiles. The app provides accessible, data-driven suspension tuning that helps riders set up their bikes safely and confidently.
NavySync is a cross-platform mobile and web system designed to centralize communication, event management, attendance tracking, learning resources, and role-based administration for Navy Junior Reserve Officers Training Corps (NJROTC) units [1]. Built with Flutter, Supabase, PostgreSQL, and a Svelte-based administrator portal, the system addresses recurring coordination problems in student organizations: scattered event information, delayed check-ins, unclear leadership permissions, and limited parent visibility [2]. NavySync uses QR-code attendance, school-scoped data, targeted announcements, role-aware event creation, and parentlinked accounts to reduce administrative friction while keeping the experience usable for cadets, instructors, parents, and guest users. This paper presents the motivation, implementation challenges, system architecture, prototype evaluation approach, related work, and future improvements for the NavySync platform.
The literature in software engineering and related fields suggests that software quality is shaped by five fundamental and mutually interacting dimensions of form: structure, space, dynamics, intelligence, and aesthetics. However, existing theories and practices treat these dimensions unevenly, often analysing and optimising them in isolation while providing a limited explicit understanding of how their interactions jointly shape software quality. Moreover, spatial and dynamical forms remain narrowly conceived, leaving important aspects of software form insufficiently articulated. These gaps constrain holistic characterization, principled design, rigorous trade-off analysis, and systematic evaluation of software systems, particularly in high-risk and mission-critical domains. To address this, the paper sought to provide a unified and holistic account of the five dimensions, their essential relationships, and how their interactions shape software quality and inform software design. A conceptualanalytical methodology synthesizing insights from software architecture, systems theory, distributed systems, control systems, intelligent systems, human-computer interaction, and design theory was employed. The principal result is a unified multidimensional framework that treats structure, space, dynamics, intelligence, and aesthetics as co-equal and mutually shaping dimensions of software form. The paper makes three primary contributions. First, it establishes an ontological framework integrating the five dimensions into a coherent model of software form and explains how their interactions give rise to core software quality attributes. Second, it develops a hierarchical model of spatial form spanning geographical through intra-node layers, broadening prevailing conceptions of software spatiality. Third, it advances a multidimensional conception of dynamical form incorporating deterministic, stochastic, adaptive, and learning-driven variability, enabling more holistic analysis and design. A case study of a global ride-hailing platform demonstrates how interactions across the dimensions create design trade-offs invisible to single-dimensional views and how multidimensional representations support more holistic design judgement. The framework implies that software design should be approached as an inherently multidimensional and cross-disciplinary activity and opens new directions for research into multidimensional software design representation, modelling, simulation, optimization, and calculations.
This work evaluates FinGPT, a financial domain-specific language model, across six key natural language processing (NLP) tasks: Sentiment Analysis, Text Classification, Named Entity Recognition, Financial Question Answering, Text Summarization, and Stock Movement Prediction. The evaluation uses finance-specific datasets to assess FinGPT's capabilities and limitations in real-world financial applications. The results show that FinGPT performs strongly in classification tasks such as sentiment analysis and headline categorization, often achieving results comparable to GPT-4. However, its performance is significantly lower in tasks that involve reasoning and generation, such as financial question answering and summarization. Comparisons with GPT-4 and human benchmarks highlight notable performance gaps, particularly in numerical accuracy and complex reasoning. Overall, the findings indicate that while FinGPT is effective for certain structured financial tasks, it is not yet a comprehensive solution. This research provides a useful benchmark for future research and underscores the need for architectural improvements and domain-specific optimization in financial language models.
We introduce a new approach of Reinforcement Learning Application for High Frequency Trading called Quantum Reinforcement Learning as our agent learns to react on ‘quantum’ individual events in Limit Order Book – single Limit Order Book updates and single trades (and optionally single Orders if provided by Exchange). We claim that such level of learning granularity allows our agent to find optimal trading strategies by on-line modeling of Market Microstructure with a maximum rate and precision.
Pneumonia is a major respiratory infection causing significant global morbidity and mortality, especially in developing nations with inadequate medical infrastructure. Early diagnosis through chest X-ray imaging is crucial but challenging. This study developed an automated computer-aided diagnosis system using deep learning to detect pneumonia from chest X-rays. An ensemble of three pre-trained convolutional neural network models (GoogLeNet, ResNet-18, DenseNet-121) was employed, with a novel weighted average ensemble technique based on evaluation metric scores. Evaluated on two public pneumonia X-ray datasets using five-fold cross-validation, the approach achieved high accuracy (98.2%, 86.7%) and sensitivity (98.19%, 86.62%), outperforming state-of-the-art methods. With pneumonia-causing over 2.5 million annual deaths worldwide, this accurate automated model can assist radiologists in timely diagnosis, especially in resource-limited settings. Its integration into clinical decision support systems has the potential to improve pneumonia management and outcomes significantly.
Text-based comments play a crucial role in providing feedback for various industries. However, effectively filtering and categorizing this feedback based on custom context-specific criteria requires sophisticated language modeling techniques. While traditional approaches have shown effectiveness, they often require a substantial amount of data to compensate for their modeling deficiencies. In this work, we focus on highlighting the performance and limitations of prompt-free few-shot text classification using open-source pre-trained sentence transformers. On the one hand, our research includes a comprehensive study across different benchmark datasets, encompassing 9 dimensions such as sentiment analysis, topic modeling, grammatical acceptance, and emotion classification. Also, we worked at making different experiences to test Prompt-Free Few-Shot Text Classification. On the other hand, we underline prompt-free few-shot classification limitations when the targeted criteria are complex. As an alternative approach, prompting an instruction-fine-tuned language model has demonstrated favorable outcomes, as proven by our application in the specific use case of “Identifying and extracting resolution results and actions from explanatory notesâ€, achieving an accuracy rate of 80%.
In a world where connecting and working with people from different countries is more and more important, the language barriers are often the main reasons why the cross border communication and collaboration is not successful. This research paper is about the use of Generative AI models, most notably the MarianMT Model and T5 Model, that enable to go through the linguistic boundaries and create the multilingual forms. The paper,on the other hand, explores the real-life application of these models in a Python environment through the Hugging Face Transformers Library. The paper goes into detailed code sample to show how these models can be used to brightly transfer textual data from one language to another apart from currently utilized models . The experimental design concerns with the translation of different sample data, this data contains individual attributes like name, age, height, weight, and the medical problems, into a number of target languages. Besides, this study not only shows the technical difficulties of model initialization and translation but also it emphasizes the wider meaning of such technology for developing cross-cultural understanding and making the world communication easier. The results underline Generative AI's potentiality to overcome language obstacles, thus enabling the worldwide cooperation, knowledge spread, and cultural exchange.
