International Journal of Emerging Multidisciplinaries: Computer Science & Artificial Intelligence

AI-Augmented Convolutional Neural Networks for Image Recognition

2026-06-27T08:24:31+00:00

The Convolutional Neural Network (CNN) is now the backbone of image recognition technology, powering computers to make accurate classifications and understand visual content. In this paper, we do a detailed analysis of the architecture of three popular CNN models namely ResNet, GoogLeNet (Inception v1) and VGGNet, and discuss the forward propagation, loss functions and back propagation mechanisms. We also explain in detail a practical implementation of ResNet50 on the CIFAR-10 dataset, which involves data preparation, building of the model, optimization of training using the Adam optimizer, and evaluation. In addition to this classic analysis, we discuss the latest developments in artificial intelligence paradigms such as Vision Transformers (ViTs), self-supervised learning, neural architecture search (NAS), and multimodal foundation models, which are transforming the image recognition field and supplementing what was already achieved by CNN models.

5G-Enabled Drone System

2026-04-21T06:50:08+00:00

This study provides an evidence-based decision framework and a quantitative feasibility assessment of a 5G-based drone system in precision agriculture. Uplink transmission times under 4G and several 5G scenarios were simulated using a data-driven model based on actual drone image transmissions (0.574 GB per mission). The simulation showed a more than 90% increase in performance using 5G, which successfully addresses the 4G transmission constraint by reducing the average transmission delay (τ1) from 1,205 seconds (20.1 minutes) under 4G to between 61 and 96 seconds (1-1.6 minutes) under 5G. However, a rural edge coverage stress test (100 simulations) showed a high standard deviation in τ1 (approximately 43 seconds), which resulted in a long tail in the latency distribution, quantifying the unreliability of the network as a primary business risk. A technical artefact, the MNDVI Heatmap algorithm, validated the system’s ability to extract valuable information from low-cost RGB sensors. To address the high capital costs, a conditional adoption approach using a Hybrid Network Architecture (5G as a service for bandwidth, a mesh network for C2 reliability) and a Drone as a Service (DaaS) business model is proposed.

An Edge-to-Access Secure Framework for Privacy-Aware Urban Surveillance

2026-07-17T05:32:07+00:00

As smart cities increasingly rely on IoT-enabled surveillance for public safety, the industry-standard practice of streaming raw footage to centralized cloud servers has introduced critical vulnerabilities regarding data privacy and accountability. Current "collect-first, protect-later" architectures create massive targets for cyberattacks and unauthorized administrative access. This report will propose an exceptional, three-layer security framework which consists of the Edge Layer, the Cloud Layer, and the Access Layer. Our particular design is to introduce a "Privacy-by-Design" methodology. At the Edge Layer, we implement an atomic, real-time anonymization process using lightweight deep learning models like YOLO (You Only Look Once), BlazeFace, and Reversible Chaotic Masking. This ensures that Personally Identifiable Information (PII) is redacted in ephemeral memory before network transmission, effectively neutralizing Man-in-the-Middle attacks. The Cloud Layer secures data that is not being transmitted via AES-256-GCM and ensures model integrity through OpenSSF Model Signing. Crucially, the Access Layer addresses the risk of data breaches and unauthorized access by utilizing Role-Based Access Control (RBAC) with a blockchain-based immutable audit ledger. To address hardware constraints, like a device with outdated hardware, this system utilizes Particle Swarm Optimization (PSO) for intelligent task offloading. After extensive comparisons, we can confirm that this holistic approach offers superior privacy protection, bandwidth efficiency, and forensic non-repudiation compared to other existing centralized surveillance models.

Deep CNN Architectures for Medical and General Image Recognition

2026-06-27T08:17:30+00:00

In image recognition, using Convolutional Neural Networks (CNNs), a computer can be trained to recognize images by learning in a hierarchical way from pixel information. In this paper, three CNN landmarks, VGGNet, ResNet and GoogLeNet, are studied and analyzed in terms of forward propagation, categorical cross-entropy loss functions, and gradient flow in back-propagation. In the field of medical imaging, an example of ResNet implementation is used for classification of brain tumors in MRI images. We also delve into the ways in which current AI developments[17-19], such as Vision Transformers, self-supervised learning, neural architecture search, and multimodal foundation models, are reshaping image recognition from traditional CNN methods.

Ransomware Defence Systems

2026-07-16T08:03:18+00:00

Ransomware has quickly become one of the newest and most harmful types of cyberattacks. Traditional security programs like antivirus programs, firewalls and signature-based Endpoint Detection and Response (EDR) programs are no longer good enough, particularly in dealing with the zero-day attacks and the new types of ransomwares that evolve their behavior to evade detection. This creates a major threat to organizations, such as operational disruptions, loss of money and irreparable data breach. To overcome these shortcomings, the Hybrid Ransomware Defense Framework (HRDF) is presented in this report, a multi-layered and dynamic security framework that will provide comprehensive protection at every phase of a ransomware attack. HRDF combines AI-based anomaly detection, deception-based honeypot validation, Zero Trust security policy, and automated responses in a single, unified defense platform. Its detection layer uses a dual-model machine learning approach, which incorporates both Random Forest classification and LSTM-based behavioral analysis as its means of detection, which allows the system to detect both structural anomalies and time-based attack patterns in real time. The automation that is inherent in HRDF significantly lowers the Mean Time to Detect (MTTD) as well as the Mean Time to Respond (MTTR), hence quick containment and recovery. In general, HRDF offers end-to-end ransomware protection that is beyond conventional approaches, and it includes proactive detection, trusted prevention, and automated recovery that is smooth enough to fit today’s advanced threat environment.

CNN-Based Image Classification on the Fashion-MNIST Dataset

2026-04-21T07:25:32+00:00

The paper is an empirical study on the use of convolutional neural network (CNN)-based multi-class image classification on the Fashion-MNIST benchmark dataset. Fashion-MNIST, consisting of 70,000 grayscale images of shoes and clothing of ten classes, was chosen as a more difficult follow-up to the canonical MNIST digit recognition benchmark, with significantly higher levels of intra-class and inter-class visual similarity to real world fashion recognition problems. A sequential CNN model was created and trained with TensorFlow and Keras and featured stacked convolutional layers with ReLU activation, max-pooling to perform spatial downsampling, dropout regularization to reduce overfitting, and a fully connected softmax output layer to estimate the probability of classes. The objective function was sparse categorical cross-entropy and the Adam optimizer was used to train the model. The held-out test partition offers empirical assessment with an ultimate classification error of 88.85, which is remarkable generalization and no major indication of overfitting. The high discriminative results of the morphologically distinctive categories, such as trousers, bags, and footwear, and the systematic inter-class confusion observed within the upper-body garment category, specifically the confusion between shirts, T-shirts, and pullovers, indicate that the results are per-class, meaning that the results are influenced by morphologically distinctive categories rather than by arbitrary combinations of such categories. Such results are placed in the context of the larger deep learning literature on fashion image recognition, and the future research directions such as data augmentation, more complex architectures, and attention-based methods are discussed.

A Quantum-Resistant Secure Communication and Key Lifecycle Framework

2026-07-14T07:59:21+00:00

Quantum computing poses a threat to modern encryption algorithms such as Rivest-Shamir-Adleman (RSA) and Elliptic Curve Cryptography (ECC) with the Shor’s algorithm, cutting cracking time from millions of years to just days. Attackers commonly use an attack known as “Store Now, Decrypt Later”. This form of attack banks on the future of quantum computing. Many post-quantum algorithms have been created to counteract the imminent attack should data leak. CRYSTALS-Kyber is one of the most secure post quantum algorithms that has been approved by the NIST. Using the CRYSTALS-Kyber as a base for a cryptographic scheme, a security framework can be developed with encryption and sessions in mind. Keeping virtual sessions as limited as possible reduces the number of openings an attacker may have to attempt to steal data. Despite the quantum threat, it is likely that a robust system is designed to prevent an attack from the present and a future threat. The system proposed will integrate session key managers along with strong encryption to ensure attackers cannot easily store the data now, nor will the data that they steal be easily cracked.

Convolutional Neural Networks for Image Classification

2026-04-21T07:17:34+00:00

In this paper, an experimental design, implementation and evaluation of a regularized Convolutional Neural Network (CNN) to classify multi-label images on the Fashion-MNIST benchmark are presented and extended to a new and urgent research area: a application of CNN-based image classification to sustainable fashion and textile circular economy systems. The architecture used by the custom CNN consists of three convolutional blocks with increasing filter depth (32, 64, 128 channels), batch normalization, data augmentation (rotation, shift, zoom), dropout regularization, early stopping, and 256-unit dense classification head with categorical cross-entropy loss and Adam as the optimizer and up to 20 epochs. The model attains a test accuracy of about 90% on 10,000-sample Fashion-MNIST test set, and a Macro-average F1-score of very high discriminative power on morphologically distinct categories (Trouser, Bag, Sandal: F1[?]0.97) and poorer on the visually similar garments (Shirt: F1[?]0.74). These findings are put in perspective of the fashion sustainability crisis on the global scale: according to the estimates provided by the United Nations Environment Programme (UNEP), 92 million tonnes of textiles are dumped into landfill each year, which makes up to 8 per cent of the total global greenhouse emissions. CNN-based visual classification The identical convolutional feature extractor, which is proven on Fashion-MNIST, is already used in intelligent textile sorting systems with 93% classification accuracy in 2025 pilot line, and in post-consumer fabric recycling pipelines to sort textiles into recycling streams at 95 percent accuracy at a rate of one item/second (Tsai and Yuan, 2025). The paper builds a technical roadmap of extending the Fashion-MNIST CNN engineering capabilities to the use in the circular economy, recognizes four AI-enabled pillars of the circle economy that are supported by the literature, and suggests specific future research directions on the cross-section of computer vision, sustainability, and regulatory compliance.

Towards Behaviour-Aware Adaptive Zero Trust Architecture

2026-07-17T07:06:21+00:00

The proposed study will present an Adaptive Behaviour-Centric Zero Trust Architecture (AB-ZTA), which targets critical weaknesses of current Zero Trust systems. Although the existing industry designs offer robust identity and access-level controls, they are still prone to heavy reliance on fixed policies, environment-specific ecosystems, as well as narrow behavioural context. These silos decrease visibility in mixed environments, responsiveness to new threats, and can generally lead to user-experience friction by forcing unnecessary repeated authentication. AB-ZTA model has a dynamic and intelligence-based approach that is constantly validating users, devices, and session based on real-time behavioural analytics and adaptive trust scoring. Risk-adaptive access control of a highly granular nature is facilitated by core ingredients such as distributed Policy Decision Point, dynamic Policy Enforcement Point, machine-learning-based anomaly detection, and context-driven data shielding. Cryptographic standards (TLS 1.3, AES-256, FIDO2, JWT/JWE) are also implemented in the architecture to ensure the privacy of communication, sessions, and data at rest. The comparative analysis demonstrates that AB-ZTA improves threat detection, minimizes the lateral mobility with the help of micro-segmentation, and minimizes data disclosure with device-bound session tokens and dynamic masking. Even though there is some performance overhead on the part of continuous behavioural verification, these impacts are manageable, and they are compensated by the level of security enhancements. Altogether, AB-ZTA is a scalable and vendor-neutral and user-conscious Zero Trust architecture that can be applied to contemporary enterprise networks.

AI-Enhanced Deep Learning Architectures for Image Recognition

2026-06-27T08:28:25+00:00

Convolutional Neural Networks (CNNs) have revolutionized image recognition by organizing features hierarchically and optimizing structures. In this paper, three popular CNN architectures (VGGNet, ResNet and GoogLeNet, also known as Inception) are compared in terms of their forward propagation, loss functions and back propagation. We delve deeper than that basic exploration and discuss the impact of current Artificial Intelligence (AI) developments such as Vision Transformers, self-supervised learning, and neural architecture search on the image recognition field. We analyze the computational efficiency, gradient flow, and scalability of each model, and illustrate that ResNet's residual connections provide the best of the three worlds of depth scalability, gradient stability and training efficiency. In addition, the emerging paradigms of AI like foundation models and multimodal learning are coming together with CNN-based solutions to shape the future generation of visual understanding systems. In this video, we will introduce you to the various neural network architectures that dominate the field of image recognition.In this video, we will introduce you to the various types of neural network architectures that dominate the field of image recognition: convolutional neural networks, deep learning networks, ResNet, VGGNet, GoogLeNet, Vision Transformers, self-supervised learning, and AI.

Towards Privacy-Preserving and Explainable CNN-Based Image Classification

2026-04-21T07:09:30+00:00

The paper introduces a two-axis extension of the standard Convolutional Neural Network (CNN) image classification studies; the technical exploration is based on an experimental baseline of Fashion-MNIST and generalizes its results to two frontier research directions of pressing societal importance Federated Learning (FL) to train distributed models privately and Explainable Artificial Intelligence (XAI) to make transparent and clinically interpretable decisions. It is a three-block sequential CNN (with 32, 64, and 64 filters Convolutional layers, max-pooling, dense classification and softmax output) that is trained on the 15,000 sample Fashion-MNIST test data with Adam optimizer categorical cross-entropy in 15 epochs (resulting in 89.57 percent accuracy and macro-averaged F1-score of 0.90). This performance profile per-class, i.e., high F1 on morphologically distinct classes (Trouser: 0.98; Bag: 0.98; Sandal: 0.96) and significantly lower performance on visually confusable classes (Shirt: 0.71) is systematically studied to incentivise a Federated Learning architecture that can quickly train CNNs on decentralised and non-IID data partitions without access to raw training examples and a Gradient-weighted Class Activation Mapping (Grad-CAM) XAI addition that can make Combined, these extensions map out a technically sound research path to deploy privacy-conscious, transparent CNN classifiers in high-stakes areas of application such as clinical diagnostic imaging, federated retail AI, and image pathology across institutions.

Behavioural Biometric Authentication via Mouse Dynamics

2026-07-17T06:24:15+00:00

Authentication is an essential component to a secure system by verifying the credentials of a user before the access right was granted to access system resources legitimately. However, previous research proved that as the computer intrusion techniques continue to advance, traditional authentication methods such as passwords and one time password (OTP) gradually reveal their security weakness, and it is no longer enough to preserve the security of the system. Therefore, a more secure authentication method, behavioural biometrics, was introduced to thwart intruders and enhance the security by verifying users’ unique behavioural pattern that was difficult to mimic. This project proposes an authentication system that integrates mouse dynamics behavioural biometrics and traditional password authentication to detect suspicious login attempts. The user needs to predefine three mouse click positions as part of their user profile. The same mouse click position and sequence must be performed during authentication before proceeding to username and password entry. Successful login requires both the mouse click coordinates and user’s credentials to match the stored profile. The proposed system was experimentally evaluated by comparing the accuracy, security and performance with traditional password authentication. The mouse dynamic biometric method has an overall accuracy of 92%. Although it performs higher latency and lower throughput than password authentication, it provides stronger security and greater resistance to brute-force attack. These results show that the proposed authentication system provides a more secure and effective approach to user authentication.

AI-Driven Deep Learning for Flower Image Recognition

2026-06-27T08:21:01+00:00

Convolutional Neural Networks (CNNs) have transformed the field of computer vision, enabling computers to recognize and classify visual data with high accuracy. This paper compares the three well known CNN architectures namely VGGNet, ResNet and Inception (GoogLeNet), specifically while looking at their forward propagation, loss functions, and gradient flow mechanisms. We propose a ResNet50 based transfer learning system for flower species classification on the TensorFlow Flowers dataset with the test accuracy of 90.33%. In addition to classical methods, we examine the application of contemporary AI methods such as Vision Transformers, self-supervised learning, neural architecture search, and multimodal foundation models, which are enhancing the capabilities of image recognition and complementing the CNN based methods.

A Two-Tier Hybrid Intrusion Detection System for IoT Networks

2026-07-16T09:42:50+00:00

The rapid growth of Internet of Things (IoT) devices has made modern attacks more vulnerable to cyberattacks. Traditional signature-based Intrusion Detection Systems (IDS) are no longer enough to keep up with new and evolving threats. Although machine learning and deep learning have improved detection accuracy, many AI-driven IDS models still face major issues. They often struggle to detect zero-day attacks, produce high false-positive rates and perform poorly with imbalanced datasets. Some models are also too computationally heavy to run efficiently in real time. To address these weaknesses, this research proposes a two-tier hybrid IDS that uses a Random Forest model for quick initial detection and a Neural Network for deeper analysis of suspicious traffic. A confidence threshold of 0.8 is used to decide whether traffic should be accepted or sent for further inspection. Using the NSL-KDD dataset, the system includes preprocessing steps such as binary mapping and structured feature extraction to support both detection stages. Our comparative analysis shows that this hybrid approach can achieve better accuracy, fewer false alarms, and stronger detection of unknown attacks compared to existing Machine Learning / Deep Learning IDS methods. It is more practical for large, diverse IoT environments because it reduces computational load while maintaining strong detection capability. Overall, the proposed architecture provides a balanced and efficient solution that overcomes key limitations of existing IDS models and offers a pathway towards a more robust real-time IoT intrusion detection.

Multi-Class Image Classification on Fashion-MNIST Using a Custom CNN

2026-04-21T07:32:47+00:00

This paper presents a custom lightweight Convolutional Neural Network (CNN) designed from scratch for multi class classification on Fashion-MNIST. The architecture employs progressive convolutional blocks (32/64/128 filters), batch normalization for stable training, max/global pooling for dimensionality reduction, dropout for regularization, and data augmentation to enhance generalization. Implemented in TensorFlow/Keras and trained over 100 epochs with Adam optimization, the model totals just 111,370 parameters.

Privacy-Preserving Intrusion Detection in Smart Traffic Networks

2026-07-16T07:53:53+00:00

Smart transportation and associated systems are becoming more susceptible as they also apply to cyber threats such as Distributed Denial-of-Service (DDoS), model poisoning, node impersonation, and ransomware lateral movement using more interconnected IoT products (traffic controllers, sensors, cameras, etc.). The current centralized Intrusion Detection Systems (IDS) have structural disadvantages, namely, high latency, bandwidth overhead, privacy exposure, and single-point collision, which limit their applicability in real and safety-critical urban settings. In order to handle these gaps, this research will suggest a Federated Learning Multi-layer Intrusion Detection System (FL-IDS) that is specifically crafted to intelligent traffic infrastructures. The architecture incorporates edge-based anomaly detection, federated collaborative learning, secure aggregation, differential privacy, encrypted communication (TLS 1.3, MQTT-S, SNMPv3), and devices-integrity (Secure Boot and firmware signing). Every intersection does its local detection and transmits the encrypted model updates, which allows them to learn globally and capture the local traffic features. Detection performance, latency, and bandwidth consumption coupled with resistance to poisoning attacks were tested within a conceptual experimental framework comprising of the CICIoT2023 dataset and trafficking simulated variations in the real world. Findings indicate that FL-IDS is better at performance compared to Cloud-IDS and traditional ML-IDS, with a high detection rate of 95% and a false-positive rate of 1.8 percent along with a detection latency of 80 ms and bandwidth consumption of 2.3 MB. With the conditions of model-poisoning, the reduction in accuracy is only as high as 8 percent, proving to be highly resilient with secure aggregation and differential privacy.

Convolutional Neural Networks for Multi-Class Image Classification

2026-04-21T07:21:27+00:00

The given work is a design, implementation and evaluation of a tailored Convolutional Neural Network (CNN) that can be trained to provide multi-class image classification in ten different clothing categories based on the Fashion-MNIST benchmark data set (Xiao, Rasul, and Vollgraf, 2017). The data sets include 70,000 28x28 grayscale 28x28 pixel images (60,000 training data and 10,000 test data). In the proposed CNN architecture, three convolutional-pooling blocks each containing a filter depth of 32, 64 and 128 are used, then finally a fully connected classifier with a softmax output layer to generate class probabilities in the ten categories. The data preprocessing steps involved pixel value rescaling in [0, 1] range, reshaping tensors to meet the Conv2D layer requirements, stratified train/validation/test splitting, and sparse integer label encoding with the sparse categorical cross-entropy loss. The Adam optimizer (Kingma and Ba, 2015) was used, with validation-based callbacks, i.e. Early Stopping and ReduceLROnPlateau, to reduce overfitting and guarantee generalizable behavior. After testing on the held-out test set, the final model had test accuracy of about 92% and test loss of about 23 percent indicating high generalization to unknown data. The confusion analysis has shown that classification errors were clustered across the visually similar category of upper-body garments, in particular, shirt, T-shirt, coat, and pullover, which is also very common in the Fashion-MNIST literature and is also primarily due to the low pixel density of the dataset (Xiao et al., 2017). This conclusion is supported by the fact that a small, purposely designed CNN architecture is an effective and computationally efficient solution to this benchmark classification problem, and that the performance gaps still exist largely due to natural limits of the data sets, and not due to architecture weaknesses.

Stable Backpropagation in Deep Image Recognition

2026-06-27T08:31:53+00:00

This paper explores the properties of back propagation and gradient flow of three basic CNN architectures – VGGNet, ResNet and Inception (GoogLeNet) – in image recognition. For each architecture, we explain the details of forward propagation, the calculation of categorical cross-entropy loss and the backward pass mechanisms. Training stability and transparent gradient monitoring is established in a VGGNet implementation on CIFAR-10. In addition to classical analysis, some modern AI paradigms such as Vision Transformers, self-supervised learning, neural architecture search, and foundation models are revolutionizing the field of image recognition, and are being introduced to augment the capabilities of CNN-based approaches.

Image Classification Using Convolutional Neural Networks

2026-04-21T07:13:20+00:00

The given paper explores the issue of Convolutional Neural Network (CNN)-based image classification on the Fashion-MNIST dataset and elaborates on the findings in relation to the impacts on smart retail artificial intelligence and the process of its implementation into practice through transfer learning and edge computing. It is a custom three-block sequential CNN, which consists of progressive convolutional layers consisting of 32, 64, and 128 filters, max-pooling, fully connected dense layer of 128 neurons, and a 10-class softmax output, trained, and evaluated on 10 epochs with the Adam optimizer and sparse categorical cross-entropy loss in TensorFlow/Keras. This model attains an approximate test accuracy of 88.8 per-class F1-scores are high in categories that are morphologically distinct, like Trouser, Bag, and Sandal (F1[?] 0.97) and lowly with those that are closely similar in appearance like Shirt and T-shirt/top by the greyscale as they may be inter-classes. These empirical results are conceptually mapped into real-world implementation scenarios: automated retailing, e-commerce product tagging with AI, virtual try-on, and fashion recognition on the edge. The article also discusses transfer learning models, specifically MobileNetV2 and EfficientNetB3, as computationally efficient frameworks to be used in computing resource-limited deployment settings, comparing their parameter efficiency and inference rate with the specialized architecture. The business environment of AI-in-fashion is the global market, which is currently USD 2.23 billion with a compound annual growth rate of 39 percent (SmartDev, 2025) and is expected to increase to USD 60 billion in 2034.

An Explainable Multi-Modal Phishing Detection Framework

2026-07-17T06:51:02+00:00

Phishing is one of the most common online threats, and it is becoming harder to detect because attackers use AI nowadays, fake website designs, and new tricks. Traditional methods like simple rules or URL reputation checks are no longer enough to stop these modern attacks. This report introduces Phishing Shield AI, a phishing detection system that uses three types of analysis: text analysis with NLP, URL and domain checking, and computer vision to compare webpage visuals. Each part gives a risk score, and the system combines them to decide whether something is phishing. It also provides clear explanations by showing which words, links, or images look suspicious. Based on the literature review, gap analysis, system design, and evaluation, the multi-modal approach improves accuracy, reduces mistakes, and adapts better to new phishing methods. Overall, Phishing Shield AI is a practical and scalable solution that performs better than single-method systems and has strong potential for future improvement.