Detection Framework Research Articles

TrCL-AGS: A Universal Sequential Triple-Stage Contrastive Learning Framework for Bacterial Detection With Across-Growth-Stage Information

TrCL-AGS: A Universal Sequential Triple-Stage Contrastive Learning Framework for Bacterial Detection With Across-Growth-Stage Information

Advanced digital image forensics: A hybrid framework for copy-move forgery detection in multimedia security.

Particularly in validating image integrity, advances in digital image analysis have profoundly affected forensic investigation. The growing reliance on digital image technology can be attributed in part to the broad availability of consistent and effective image-capturing technologies. The simplicity of changing image content thanks to advanced image-editing technologies presents fresh difficulties for forensic analysis. A structured hybrid framework is presented for finding important objects in images. It does this by using fast Fourier transformation (FFT) for frequency domain filtering, scale-invariant feature transformation (SIFT), and oriented FAST and rotated BRIEF (ORB) to pull out key points. The MobilenetV2 and VGG16 models extract features from key point areas to detect copy-move forgery. After that, an attention mechanism combines and normalizes these aspects. Key point matching uses the Euclidean distance; DBSCAN clustering groups pertinent key points for object localization. The suggested approach shows better performance than current methods and detects image copy-move forgery rather successfully. The framework's robustness is verified against image blurring, contrast alteration, color reduction, image compression, and brightness change among other post-processing techniques. Since photographs are altered, traditional approaches can struggle with a lot of variety; however, the proposed method combines advanced deep learning models and clustering techniques to make detection more accurate. Extensive testing on five benchmark copy-move forgeries datasets reveals that the suggested strategy may beat present techniques. This work offers a sophisticated automated approach to guarantee digital image integrity and identify image manipulation.

Early detection of mental health disorders using machine learning models using behavioral and voice data analysis

People of all demographics are impacted by mental illness, which has become a widespread and international health problem. Effective treatment and support for mental illnesses depend on early discovery and precise diagnosis. Notably, delayed diagnosis may lead to suicidal thoughts, destructive behaviour, and death. Manual diagnosis is time-consuming and laborious. With the advent of AI, this research aims to develop a novel mental health disorder detection network with the objective of maximum accuracy and early discovery. For this reason, this study presents a novel framework for the early detection of mental illness disorders using a multi-modal approach combining speech and behavioral data. This framework preprocesses and analyzes two distinct datasets to handle missing values, normalize data, and eliminate outliers. The proposed NeuroVibeNet combines Improved Random Forest (IRF) and Light Gradient-Boosting Machine (LightGBM) for behavioral data and Hybrid Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) for voice data. Finally, a weighted voting mechanism is applied to consolidate predictions. The proposed model achieves robust performance and a competitive accuracy of 99.06% in distinguishing normal and pathological conditions. This framework validates the feasibility of multi-modal data integration for reliable and early mental illness detection.

Securing Telehealth with State-of-the -Art Machine Learning: A DevSecOps Framework for Real-Time Phishing Detection

Telehealth’s rapid expansion, driven by global health crises, has heightened its susceptibility to phishing, demanding cutting-edge countermeasures. This paper presents a pioneering DevSecOps pipeline tailored for telehealth, integrating state-of-the-art (SOTA) Long Short-Term Memory (LSTM) and Graph Neural Network (GNN) models to achieve 95% accuracy in real-time phishing URL detection. Departing from conventional hybrid ML, our approach harnesses LSTM for sequential URL pattern recognition and GNN for graph-based campaign analysis, validated by the “PhishGuard” browser plugin, which preemptively blocks malicious portals. Augmented by behavioral analytics, EHR correlation, and rigorous ablation studies, our framework surpasses traditional methods by 5-10% in recall while ensuring HIPAA compliance and scalability. This solution redefines telehealth cybersecurity with adaptive, SOTA-driven protection against evolving threats.

A Novel Algorithm for Adaptive Detection and Tracking of Extended Targets Using Millimeter-Wave Imaging Radar

A high-resolution imaging radar is exceptionally well-suited for the detection and perception of extended targets (ETs), as it provides a comprehensive representation of the spatial distribution of target scattering characteristics. In this work, we propose an adaptive detection and tracking framework for non-cooperative ETs based on radar imaging. The framework leverages the statistical properties of ETs in radar imaging to construct a target distribution model and introduces an adaptive ET detection and tracking algorithm based on Scattering Point Shift (SPS). This algorithm is designed to track ETs with internal motion characterized by multiple scattering points. The initial target distribution is estimated using two-dimensional kernel density estimation (2D-KDE). Compared to existing ET tracking algorithms, the proposed SPS method demonstrates superior universality in accommodating diverse scattering point distributions and integrates detection and tracking into a unified process, thereby significantly improving information utilization efficiency. The effectiveness of the algorithm is validated through extensive simulations and real-world data collected using a millimeter-wave (mmWave) imaging radar operating in the Linear Frequency Modulated Continuous Wave (LFMCW) mode.

UMFNet: Frequency-Guided Multi-Scale Fusion with Dynamic Noise Suppression for Robust Low-Light Object Detection

The dominant low-light object detectors face the following spectral trilemma: (1) the loss of high-frequency structural details during denoising, (2) the amplification of low-frequency illumination distortion, and (3) cross-band interference in multi-scale features. To resolve these intertwined challenges, we present UMFNet—a frequency-guided detection framework that unifies adaptive frequency distillation with inter-band attention coordination. Our technical breakthroughs manifest through three key innovations: (1) a frequency-adaptive fusion (FAF) module employing learnable wavelet kernels (16–64 decomposition basis) with dynamic SNR-gated thresholding, achieving an 89.7% photon utilization rate in ≤1 lux conditions—2.4× higher than fixed-basis approaches; (2) a spatial-channel coordinated attention (SCCA) mechanism with dual-domain nonlinear gating that reduces high-frequency hallucination by 37% through parametric phase alignment (verified via gradient direction alignment coefficient ρG = 0.93); (3) a spectral perception loss combining the frequency-weighted structural similarity index measure (SSIM) with gradient-aware focal modulation, enforcing physics-constrained feature recovery. Extensive validation demonstrates UMFNet’s leadership: 73.1% mAP@50 on EXDark (+6.4% over YOLOv8 baseline), 58.7% on DarkFace (+3.1% over GLARE), and 40.2% on thermal FLIR ADAS (+9.7% improvement). This work pioneers a new paradigm for precision-critical vision systems in photon-starved environments.

UAV-based water pollutants detection and classification framework using multi-modal and multi-sensor ensemble learning.

The massive increment in water pollutants due to the release of plastic, industrial, and household waste has threatened the delicate balance of ecosystems and the well-being of human life. Therefore, detection and monitoring of such water pollutants have become an essential task for the widespread and open surface water bodies. Recent advancements in UAVs with Computer Vision (CV) models and communication technologies have given the scope to automate the process of pollutant monitoring in such surface water bodies, minimizing human intervention. This paper presents a comprehensive framework integrating UAVs for autonomous data collection and pollutant classification. The customized YOLOv5 model is utilized for both the classification and detection of water pollutants, enhancing efficiency and accuracy. Moreover, we propose a multi-modal feature extraction module that uses Vision Transformer (ViT), YOLOv5, and NodeMCU sensors to create a comprehensive data representation. The extracted features are then classified using an ensemble model combining TabNet and XGBoost, improving the overall classification performance. An image dataset for water pollutant detection has been prepared using video sequences captured by a UAV-based camera at different zoom levels and altitudes. The results show that the proposed model performed better than the MobileNet, YOLOv4, YOLOv5s, and YOLOv8 in terms of both the response time and the mAP of ( ) for Algae, ( ) for trash, and ( ) for the classification of pollutants of multi-classes. This work aims to advance the deployment of UAVs for environmental monitoring, providing an efficient and scalable solution for water pollutant detection.

Deep Learning Based Phishing Detection System

Malicious URLs and websites continue to undermine online security, with search engines inadvertently becoming platforms for fraudulent sites. Traditional phishing detection methods often based on white lists, blacklists, or single-model approaches fail to address the evolving sophistication of phishing attacks. This persistent threat highlights the urgent need for more advanced and adaptive security measures that can reliably identify unsafe URLs in real time. Motivated by these challenges, our research introduces an enhanced phishing detection framework that integrates Natural Language Processing (NLP) with a combination of machine learning algorithms specifically, Support Vector Machine (SVM), Random Forest, and Decision Tree. The SVM algorithm is chosen for its robustness in handling high-dimensional data, while Random Forest and Decision Tree contribute through ensemble learning and interpretability, respectively. Together, these methods form a comprehensive system that accurately differentiates between malicious and legitimate URLs. Additionally, the system incorporates AES encryption to secure sensitive user data, ensuring that browsing history and other critical information remain confidential. The significance of this research lies in its dual contribution to improving cybersecurity and safeguarding user privacy. By effectively detecting phishing attempts and encrypting user data, our approach not only mitigates the risks associated with malicious URLs but also establishes a higher standard for protecting sensitive information online. This integrated solution paves the way for more resilient defences against phishing attacks, offering users enhanced security without compromising privacy.

SCL-YOLOv8n based rice disease lightweight detection method

Abstract To address the challenges posed by complex rice disease features, low detection accuracy, and large model size, this paper, we propose slim cross-level lightweight YOLOv8n (SCL-YOLOv8n), an enhanced lightweight target detection framework based on YOLOv8n. Firstly, a novel slim-neck network architecture was designed to optimize concatenation of feature representations, thereby reducing computational cost and the number of parameters. Secondly, the receptive-field collaborative attention cross-stage partial network (RFCA-CSP) was proposed, integrating convolutional neural networks with the transformer architecture to enhance feature extraction capabilities while minimizing computational overhead. Finally, the lightweight shared-convolution with separated batch normalization and dynamic anchors (LSCSBD) detection head was incorporated to enhance the model’s computational efficiency through the implementation of techniques including shared convolution, separated batch normalization, and dynamic anchor generation. Experimental results demonstrate that the improved SCL-YOLOv8n increased the mAP50 by 5.0%. points compared with the traditional YOLOv8n. Concurrently, it decreased the parameter count to 1.93 M and the computational volume to 5.5 GFLOPs. These represent reductions of 35.7% and 31.3% respectively when compared with the original model. The SCL-YOLOv8n architecture exhibits dual advantages, it not only enhances the accuracy of object detection but also achieves substantial reductions in both the number of parameters and computational complexity. This advancement offers an effective approach for detecting rice diseases in complex backgrounds, thereby demonstrating significant potential for application in agricultural disease monitoring scenarios.

A Cybersecurity Detection Platform Integrating IOTA DLT and IPFS for Vulnerability Management

In response to the Cybersecurity Law, organizations face numerous management and technical requirements. Detection techniques such as vulnerability scanning and penetration testing are employed to identify risks. Addressing these vulnerabilities demands substantial manpower, time, and financial resources. Security concerns also arise during digital file transmission and remediation efforts. This study proposes a security detection platform with step-by-step implementation guidelines, enabling resource-limited units to replicate the setup and address security gaps. It compares detection results between open-source and commercial tools, highlighting key differences and offering remediation strategies. Numerous digital files (e.g., test reports) are generated during testing. To ensure secure storage and sharing, the system integrates IOTA’s distributed ledger and IPFS, generating HASH values and uploading files on-chain to preserve integrity and authenticity. The objective is to deliver a scalable, cost-effective security detection framework that enhances system resilience while minimizing resource consumption.

An adaptive hybrid framework for IIoT intrusion detection using neural networks and feature optimization using genetic algorithms

In Industrial Internet of Things (IIoT) networks, securing device connectivity through effective intrusion detection systems is essential for maintaining operational integrity. This paper presents an adaptive hybrid framework for IIoT intrusion detection that combines Artificial Neural Networks (ANNs) with Genetic Algorithms (GA) for feature optimization. This study utilizes the dataset, which is widely recognized benchmark for intrusion detection research. The dataset comprises 625783 network traffic samples, classified into five categories: Denial-of-Service (DoS), Probe, Remote-to-Local (R2L), User-to-Root (U2R), and Normal traffic. Initially, an ANN model was developed, yielding high accuracy but exhibiting signs of overfitting. To enhance generalization, we introduced L2 regularization, adjusted dropout rates, and optimized the learning rate, ultimately achieving a 99.7% validation accuracy with an AUC score of 0.9969. Additionally, Genetic Algorithms were employed to optimize feature selection, further refining the ANN’s input space to improve computational efficiency without sacrificing predictive power. After training over 50 epochs with early stopping, the model demonstrated exceptional robustness, achieving 99.5% accuracy on the test set, with precision and recall values of 0.97 and 0.98, respectively. This combination of ANN and GA yielded a highly efficient and sensitive framework, providing enhanced detection of anomalies in IIoT environments. The proposed hybrid model thus establishes a robust solution for real-time IIoT security, outperforming conventional detection systems through a strategic blend of neural learning and evolutionary optimization.

Toward the detection of intracranial hemorrhage: a transfer learning approach

Traumatic brain injury can cause intracranial hemorrhages, which, if not diagnosed and treated early, may lead to fatal outcomes due to excessive bleeding inside the cranium. In the United States, stroke is the fifth-leading cause of death, and approximately 10% of strokes result from intracranial hemorrhages. Identifying the presence, location, and type of hemorrhage is a critical step in treating emergency room patients. To determine the position and size of hemorrhages, X-ray computed tomography (CT) scans are commonly employed. While radiologists are highly skilled in analyzing CT scan images, the process is time-consuming. This study deals with intracranial hemorrhage detection using a deep-learning approach. First, we introduce convolutional neural networks (CNNs), a type of neural network designed for image-based datasets. Subsequently, we discuss various hyperparameter optimization techniques to enhance CNN training efficiency. As CNN training can be computationally expensive and time-intensive in many instances, we address this challenge by leveraging transfer learning with pre-trained models. We explore different transfer learning architectures, including VGGNet, AlexNet, EfficientNetB2, ResNet, MobileNet, and InceptionNet. The proposed framework for intracranial hemorrhage detection is implemented using transfer learning on the Radiological Society of North America Intracranial Hemorrhage Detection dataset, which is publicly available on Kaggle. Specifically, VGGNet is employed within this framework using powerful deep-learning libraries such as TensorFlow. This methodology can also be generalized to classify CT scan images in other biomedical domains.

The ADBPSO-LightGBM internal threat detection framework based on hybrid data balancing

The misuse of privileges by users can lead to significant reputational and financial losses for enterprises. To reduce the risk of information leakage, it is crucial to detect and analyze abnormal behaviours of internal employees. Firstly, based on the characteristics of internal employee behaviour, a data filter strategy based on user behaviour is proposed. Then, a data balancing strategy based on the concept of hybrid sampling is introduced. Moreover, to further construct the behaviour model, an improved particle swarm optimization algorithm based on adaptive delay and genetic factors is proposed, and it is used to search for the optimal parameters of LightGBM. Experimental results demonstrate that the proposed method is highly effective in detecting internal threats.

RDLNet: A Channel pruning-based traffic object detection algorithm

Abstract Accurate traffic object detection plays a critical role in intelligent transportation systems, yet it faces challenges in balancing real-time performance with detection accuracy. This challenge persists particularly under complex scenarios involving occlusions, multi-scale objects, and dynamic environmental changes. To address these challenges, this paper introduces RDLNet (RepViT backbone with a DBB-Gelan-based Neck (DG Neck) and Layer-Adaptive Magnitude-based Pruning (LAMP) Network), a lightweight traffic detection framework that synergistically integrates adaptive channel pruning and multi-scale feature refinement. The proposed architecture achieves enhanced model efficiency while preserving detection accuracy through hierarchical feature fusion mechanisms. The conventional backbone is replaced by RepViT, which enables the capture of multi-scale spatial dependencies. A dynamic feature fusion neck network called DG-Neck, composed of stacked DG modules, is introduced to adaptively aggregate features from different layers and enhance multi-scale feature discrimination. Additionally, Layer-Adaptive Magnitude-based Pruning is applied to compress the model by selectively removing redundant channels while preserving essential detection pathways. Comprehensive experimental evaluations confirm RDLNet’s superiority over state-of-the-art methods. On the DAWN dataset, RDLNet demonstrates significant advantages over YOLOv8n by achieving 39.5% fewer parameters and reducing FLOPs by 14.8%. It also delivers substantial detection improvements, with 43.5% higher precision, 23.1% increased recall, and gains of 14.0% and 9.0% in mAP50 and mAP50-95, ultimately reaching 0.366 and 0.205. Comparative experiments with advanced models consistently reveal lower parameter counts and enhanced detection performance. Furthermore, its robustness has been validated on the VOC2012 dataset. RDLNet’s lightweight design and hierarchical feature learning provide a deployable solution for resource-constrained edge environments, thereby advancing real-time traffic monitoring systems through optimized accuracy-efficiency trade-offs.

A Novel Framework for Malware Detection Using Entropy-Based Statistical Features and Machine Learning Models Across File Types

Abstract As cyber threats continue to evolve, the accurate detection of malicious files has
become increasingly crucial. Traditional approaches often fall short due to limited
adaptability to diverse file types and a high incidence of false predictions. This study
addresses these gaps by systematically evaluating entropy-based features in conjunction
with machine learning (ML) models for malicious file detection. Using diverse file
types—documents, images, and compressed files—we employed byte-level analysis of
each file’s raw stream—without any tokenization into chunks—to compute Shannon
entropy (F1) and R´enyi entropy (α = 2, 4, 6) (F2), along with statistical measures
(25th percentile, mean, and 75th percentile of F1+F2) as key features. From these
streams, we extracted statistical measures—25th percentile, mean, and 75th percentile
of F1 and F2—which capture fine-grained entropy variations indicative of potential
malicious patterns. Files exceeding an entropy threshold of 7.95 were classified as
malware-infected, with the NapierOne dataset serving as the basis for analysis. To
minimize false negatives, we applied three popular ML models—Random Forest (RF),
Decision Tree (DT), and Naive Bayes (NB). Our findings reveal that image-based files
are particularly vulnerable to malware, while Shannon entropy combined with RF and
DT models effectively reduced false negative predictions. The RF model emerged as
the best-performing approach, achieving over 99% detection accuracy across all file
types. This research provides a novel integration of entropy-based features with ML
models, demonstrating their synergy for improving malware detection. By addressing
key limitations in previous approaches, this study lays the foundation for scalable,
accurate, and robust detection frameworks, paving the way for future advancements
in combating dynamic cybersecurity threats.

GFADE: generalized feature adaptation and discrimination enhancement for deepfake detection

With the rapid advancement of deep generative techniques, such as generative adversarial networks (GANs), the creation of realistic fake images and videos has become increasingly accessible, raising significant security and privacy concerns. Although existing deepfake detection methods perform well within a single dataset, they often experience substantial performance degradation when applied across datasets or manipulation types. To address this challenge, we propose a novel deepfake detection framework that combines multiple loss functions and the MixStyle technique. By integrating Cross-Entropy Loss, ArcFace loss, and Focal Loss, our model enhances its discriminative power to better handle complex forgery characteristics and effectively mitigate data imbalance. Additionally, the MixStyle technique introduces diverse visual styles during training, further improving the model’s generalization across different datasets and manipulation scenarios. Experimental results demonstrate that our method achieves superior detection accuracy across a range of cross-dataset and cross-manipulation tests, significantly improving model robustness and generalizability.

A Comprehensive Framework for Frame Detection Leveraging SIFT and Visual Feature Characterization

This project focuses on developing a system that can identify videos using individual frames or short sequences. This is a complex task, but it has the potential to revolutionize how we interact with video content in many industries, from entertainment to security. The ability to identify videos from just a still frame or short video segment is a complex yet highly demanded task in industries ranging from entertainment to security. The system will use visual feature extraction and a comprehensive database to match frames to videos. The methodology involves using a combination of SIFT, YOLOv5, and ResNet-50 to process and analyze the frames. ChromaDB, a vector database for AI applications, is used to store and search for matches. The system will then use a modified ensemble ranking system that considers factors like frequency, consistency, and tag coverage to calculate a confidence score for each match. This score will be displayed to the user along with the matched videos. The project aims to provide a user-friendly interface that allows users to upload images and view the predicted videos, as well as the calculations performed during the matching process. Future improvements include refining the algorithm for finding unique frames, enhancing the user interface with history tracking, and improving the confidence calculation algorithm.

AI-Driven Fraud Detection: Enhancing Claims Analytics with Real-Time Streaming and Behavioral Biometrics

Scammers and frauds have always been challenging and at the same time the most undetected threat to the insurer world. We will present the experimental work we have done on our novel pattern discovery framework for fraud detection, which is based on our examination of a historical claims data repository in ClaimCenter, together with real-time data mining techniques to detect emergent fraudulent trends. AI also improves existing rule-based engines by allowing them to incorporate anomaly detection into their existing systems, and to actively respond in order to prevent future fraud attacks. Using graph databases such as Neo4j, you can perform an in-depth analysis of relationships between claims, policyholders, and external actors, uncovering hidden links that could indicate potential fraud. The paper also explores behavioral biometrics and pattern analysis to provide an addition-evolving user behavior profiling as unique, creating an additional security layer that provides a better tool in the toolbox for prevention of fraud. They also explore real-time streaming analytics platforms such as Apache Flink and Apache Spark for continuous monitoring, real-time detection of fraudulent activities, reduced latency, and improved response times. Through this paper, we introduce a holistic framework that embraces these technological developments, demonstrating how they can effectively enhance fraud detection, expedite claim managing, and reduce financial risks. Ultimately, these solutions will democratize a proactive, data-centered approach to addressing fraud, allowing insurers to remain one-third of the way ahead of fraudsters in a particularly complex and fast-moving data world.

AI-powered Alzheimer’s diagnosis: Integrating cognitive monitoring, IoT, and secure edge computing

This study proposes a privacy-preserving, multi-modal AI framework for the early detection of Alzheimer’s disease (AD), addressing the limitations of conventional single-modal diagnostic systems. The model fuses heterogeneous data sources, including physiological signals from wearable IoT devices, neuroimaging biomarkers extracted from T1-weighted MRI scans, and environmental context derived from smart home sensors. A hybrid architecture incorporating temporal CNN-LSTM networks, 3D ResNet models, attention layers, and graph neural networks is employed to extract and integrate cross-modal features. Federated learning with differential privacy (ε = 1.0) enables secure and decentralized training across distributed healthcare nodes, ensuring compliance with HIPAA and GDPR. Experimental validation on real-world datasets such as ADNI-4 and IoT-HOME shows a diagnostic accuracy of 97.3%, with a 12% improvement in recall over single-modality baselines. The system achieves sub-150 millisecond inference latency on resource-constrained edge devices through quantization and kernel pruning. Results demonstrate robust convergence, high interpretability via SHAP explanations, and scalability in heterogeneous clinical environments. The framework offers a technically robust, ethically aligned, and practically deployable solution for real-time, edge-enabled Alzheimer’s monitoring in both institutional and home-care settings.

MC-ASFF-ShipYOLO: Improved Algorithm for Small-Target and Multi-Scale Ship Detection for Synthetic Aperture Radar (SAR) Images.

Synthetic aperture radar (SAR) ship detection holds significant application value in maritime monitoring, marine traffic management, and safety maintenance. Despite remarkable advances in deep-learning-based detection methods, performance remains constrained by the vast size differences between ships, limited feature information of small targets, and complex environmental interference in SAR imagery. Although many studies have separately tackled small target identification and multi-scale detection in SAR imagery, integrated approaches that jointly address both challenges within a unified framework for SAR ship detection are still relatively scarce. This study presents MC-ASFF-ShipYOLO (Monte Carlo Attention-Adaptively Spatial Feature Fusion-ShipYOLO), a novel framework addressing both small target recognition and multi-scale ship detection challenges. Two key innovations distinguish our approach: (1) We introduce a Monte Carlo Attention (MCAttn) module into the backbone network that employs random sampling pooling operations to generate attention maps for feature map weighting, enhancing focus on small targets and improving their detection performance. (2) We add Adaptively Spatial Feature Fusion (ASFF) modules to the detection head that adaptively learn spatial fusion weights across feature layers and perform dynamic feature fusion, ensuring consistent ship representations across scales and mitigating feature conflicts, thereby enhancing multi-scale detection capability. Experiments are conducted on a newly constructed dataset combining HRSID and SSDD. Ablation experiment results demonstrate that, compared to the baseline, MC-ASFF-ShipYOLO achieves improvements of 1.39% in precision, 2.63% in recall, 2.28% in AP50, and 3.04% in AP, indicating a significant enhancement in overall detection performance. Furthermore, comparative experiments show that our method outperforms mainstream models. Even under high-confidence thresholds, MC-ASFF-ShipYOLO is capable of predicting more high-quality detection boxes, offering a valuable solution for advancing SAR ship detection technology.