Optimizing Neural Network Architectures with TensorFlow and Keras for Scalable Deep Learning

Muna; Bushra; Dheyauldeen Ahmed; Oday Ali

doi:https://doi.org/10.54216/JISIoT.180108

Fusion: Practice and Applications | 2026

Secure and Decentralized Plant Disease Detection via Federated Learning with Differential Privacy and Homomorphic Encryption

Plant disease detection using deep learning has achieved high accuracy, but traditional centralized training poses significant privacy risks and incurs high data transmission costs. This study presents a privacy-preserving federated learning (FL) framework for plant disease diagnosis that enables decentralized model training across geographically distributed agricultural sites. Rather than transferring raw farm data to a central server, local models are trained on edge devices and share only model updates. To address data heterogeneity from diverse climates, soils, and plant species, we introduce adaptive aggregation strategies that improve model generalization. Furthermore, we incorporate differential privacy and homomorphic encryption to ensure secure model updates and protect sensitive information from potential breaches. Experimental evaluations on benchmark datasets, including Plant Village and real-world field images, show that the proposed FL-based system achieves comparable accuracy to centralized models while significantly enhancing data privacy and reducing communication overhead. The framework maintains over 93% classification accuracy across 38 plant disease categories, with minimal degradation from added privacy mechanisms. Additionally, we analyze the trade-off between accuracy and communication efficiency, demonstrating the method’s practicality in bandwidth-constrained rural environments. The proposed system offers a scalable, secure, and field-deployable solution for real-time plant disease monitoring, supporting the widespread adoption of AI in precision agriculture without compromising data confidentiality.

groups

Vetripriya M. mail -

S. Amsavalli mail -

R. Sivasankari mail -

Vetri Selvan M. mail -

N. Kanimozhi mail

link https://doi.org/10.54216/FPA.210113

Volume & Issue

Vol. Volume 21 / Iss. Issue 1

Details open_in_new

Fusion: Practice and Applications | 2026

Novel Prediction on Breast Cancer through Lazy Learning Approach by Linear Neural Network Search with Distance with Euclidean

Breast cancer is the most prevalent cancer-affecting women worldwide and remains a major cause of mortality. Early detection and accurate prognosis are critical to improving survival outcomes. This study introduces a novel predictive model for breast cancer diagnosis that integrates a lazy learning paradigm with the K-Nearest Neighbors (KNN) algorithm, optimized through a Linear Nearest Neighbor (NN) Search technique and the use of Euclidean distance as the similarity measure. The dataset, comprising 4,024 patient records with 15 clinical and demographic attributes, was obtained from a public repository and underwent rigorous preprocessing, including handling of missing values, normalization, and categorical encoding. The classification model was trained and evaluated using 1:9 cross-validation, with K values ranging from 1 to 9 and a constant batch size of 100 to identify the optimal configuration. Among various configurations tested, the model with K=5 demonstrated the highest performance, achieving an accuracy of 88.02%, precision of 0.87, and recall of 0.88. Additional performance metrics such as F-measure, Matthews Correlation Coefficient (MCC), and Kappa statistic further confirmed the robustness of the selected configuration. The proposed model shows superior predictive capability compared to traditional settings and can serve as a decision-support tool for clinicians. The findings suggest that the combination of lazy learning, effective neighbor search strategy, and robust distance metric can substantially enhance the predictive accuracy of breast cancer diagnosis. This study highlights the potential of machine learning-based tools in clinical oncology, offering a data-driven approach for early intervention and patient outcome improvement.

groups

S. Amsavalli mail -

Vetripriya M. mail -

R. Sivasankari mail -

Vetri Selvan M. mail -

Vijayakumar K. mail

link https://doi.org/10.54216/FPA.210114

Volume & Issue

Vol. Volume 21 / Iss. Issue 1

Details open_in_new

International Journal of Neutrosophic Science | 2026

Neutrosophic of γ-BCK -Algebra

The most important applications of an algebra like BCK-Algebra. As a generalization of ring, we study γ- semi-ring and γ-ring in invarianent neutrosophic set. Neutrosophic concepts are widely used in the field of mathematics and other sciences, especially in studying the Algebra. In this paper, we present the concept of neutrosophic γ-BCK-Algebras as an example of this generalization. We also present neutrosophic sub-algebra, neutrosophic ideal and some other type structure algebraic. We proved that if f : AI → N I is a homomorphism of neutrosophic γ-BCK-algebras AI and NI, then f is injective if and only if neutrosophic ker(f ) = {0I}. Also, we presented, if NI be a normal neutrosophic subalgebra of neutrosophic γ-BCK- algebra AI, then ” &sim; N I ” is a congruence relation.

groups

Dunia Alawi Jarwan mail -

Amenah Hassan Ibrahim mail -

Majid Mohammed Abed mail

link https://doi.org/10.54216/IJNS.270102

Volume & Issue

Vol. Volume 27 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

ChainGuard 6G+: A Secure and Private Architecture for Wireless Communication Using Federated Learning and Blockchain in IoT Networks

The advent of 6G wireless communication systems and the widespread proliferation of Internet of Things devices have necessitated advanced frameworks for secure, private, and intelligent data management. ChainGuard 6G+, a novel privacy-preserving architecture, which integrates Federated Learning with Blockchain, is introduced in this paper to offer data security, integrity, and anomaly detection features for IoT-enabled 6G networks. FL facilitates decentralized model training across distributed edge nodes, thus keeping local data on-device with model updates shared. This ensures user privacy, particularly valuable in sensitive applications such as healthcare, financial services, and industrial IoT networks. For further strengthening privacy, Differential Privacy is applied by introducing statistical noise into model updates, masking individual contributions without degrading learning accuracy. Blockchain is incorporated as an immutable ledger to record model parameters and training securely, enabling traceability and tamper-evident model provenance. Role-based access control for secure data and model access, end-to-end encryption, and secure transmission protocols are included in the architecture. Experimental results demonstrate the efficacy of the system under consideration using a 6G Network Slice Security Attack Detection Dataset, with synthetic and real attacks on various network slices. Performance evaluation reveals that ChainGuard 6G+ not only ensures data privacy but also has excellent detection rates against DoS, DDoS, and spoofing attacks. The proposed framework achieves an overall attack detection accuracy of 99.1%, implemented and experimented using Python, revealing its promise as a secure, scalable solution for future wireless secure communication networks.    

groups

Saleh Ali Alomari mail

link https://doi.org/10.54216/JISIoT.180102

Volume & Issue

Vol. Volume 18 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Enhancing Intrusion Detection System Transparency Using SHAP-Driven Support Vector Machine Tuned by Harris Hawks Optimization

Due to the increasing prevalence of network attacks, maintaining network security has become significantly more challenging. An Intrusion Detection System (IDS) is a critical tool for addressing security vulnerabilities. IDSs play a vital role in monitoring network traffic and identifying malicious activities. However, two major challenges hinder IDS performance: data imbalance, which weakens the detection of minority class attacks, and overfitting in traditional classifiers such as Support Vector Machines (SVM). This study proposes a novel and transparent IDS framework that integrates several advanced techniques: Variational Autoencoder (VAE) for data augmentation, Mutual Information-based feature selection, Harris Hawks Optimization (HHO) for hyperparameter tuning of the SVM, and SHAP (SHapley Additive exPlanations) for interpretability. VAE is utilized to generate synthetic instances for minority classes, effectively addressing class imbalance. Feature selection is employed to reduce dimensionality and enhance generalization performance. The HHO algorithm is used to adaptively tune the hyperparameters of the SVM, thereby optimizing classification accuracy while mitigating overfitting. Finally, SHAP values are employed to interpret the SVM’s decisions, enhancing the transparency and trustworthiness of the system. Experimental evaluations conducted on two benchmark IDS datasets, UNSW-NB15 and NSL-KDD, demonstrate that the proposed VAE-HHO-SVM framework outperforms existing models in terms of accuracy, robustness, and interpretability. The results confirm the effectiveness of combining optimization, explainable AI, and data balancing strategies in modern IDS development. Specifically, the proposed method achieves an accuracy of 98.42% on the NSL-KDD dataset and 97.45% on the UNSW-NB15 dataset—an improvement of 3.17% over other methods.

groups

Noor Flayyih Hasan mail

link https://doi.org/10.54216/JISIoT.180103

Volume & Issue

Vol. Volume 18 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Blade Server Attack Detection and Mitigation Framework in Cloud Computing Using SSGRU and GGSSO

Malicious activities that seek to disrupt cloud communication are cybersecurity threats. Nevertheless, none of the existing works focused on detecting the attacks that happened in the Blade Server (BS) in the cloud. Therefore, this paper proposes an efficient Intrusion Detection System (IDS) framework for BS in the cloud by utilizing Kerberos-based Exponential Mestre-Brainstrass Curve Cryptography (KEMBCC) and Sechsoftwave and Sparsele-centric Gated Recurrent Unit (SSGRU). Primarily, the cloud users are registered into the network. Then, the incoming data are encrypted. Here, to balance the incoming loads, BS is used. To detect attacks in BS, IDS is implemented. Initially, the data are preprocessed. Further, the big data are handled in the IDS. Afterward, the features are extracted and optimal features are chosen from it. Thereafter, to classify the attack and normal BS, the SSGRU classifier is used. After that, by generating a Sankey diagram, the attacked and non-attacked blades in the BS are differentiated. Next, the attacked blades are isolated, whereas the non-attacked blades are further used for load balancing on the cloud. According to the analysis results, this model performed superior to the other models by attaining an accuracy of 99.43%.

groups

Waleed Kh. Hussein mail -

Ghaith J. Mohammed mail -

Ahmed Salih Al-Obaidi mail -

Massila Kamalrudin mail -

Mustafa Musa mail

link https://doi.org/10.54216/JISIoT.180104

Volume & Issue

Vol. Volume 18 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Optimized Deep Learning Models for Forecasting Evaporation in Almaty Using Gray Wolf Optimization

The reliable estimation of evaporation is essential for proper water resource planning, particularly in scenarios governed by climatic variability. This work proposes the application of advanced deep learning methods—namely Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU)—optimized by the Gray Wolf Optimization (GWO) algorithm in predicting monthly evaporation values over Almaty, Kazakhstan. Furthermore, the models were optimized for best performance through the adjustment of key hyperparameters such as the number of hidden units, dropout rates, and learning rates. Among candidate models for evaluation, the optimal model with smallest MSE (0.6162) and maximum value of R-squared (0.9335) was LSTM-GWO, indicating strong correlation with actual values. Performance measures such as RMSE, MAE, and MAPE strongly indicated the improved generalization strength of LSTM-GWO compared to BiLSTM and GRU. Forecasts for 2023 indicated seasonal patterns persistently expressed as maximum evaporation during summer seasons. The results detail the potential of deep learning algorithms tuned to improve the precision of hydrological forecasting specifically for semi-arid areas.

groups

Ruaa Azzah Suhail mail -

Osama Salim Hameed mail -

El-Sayed M. El-Kenawy mail -

Marwa M. Eid mail

link https://doi.org/10.54216/JISIoT.180105

Volume & Issue

Vol. Volume 18 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Keystroke Dynamics System for User Authentication Using SVM Classifier

As people increasingly rely on computers to store sensitive information and interact with various technologies, the need for low-cost, effective security measures has become more critical than ever. One such method is keystroke dynamics, which analyzes a person’s typing rhythm on digital devices. This behavioral biometric approach enhances the security and reliability of user authentication systems and contributes to improved cybersecurity. This study aims to reduce authentication risks by encouraging the adoption of keystroke-based verification methods. The research uses a fixed-text password dataset (.tie5Roanl), collected from 51 users who typed the password over eight sessions conducted on alternating days, capturing variations in mood and typing behavior. Seven models were developed, each following a structured seven-phase process. The first phase involved loading the CMU Keystroke Dynamics Benchmark dataset. The second focused on data preprocessing. In the third phase, new keystroke features were engineered from the original dataset. The fourth phase involved feature selection across various types: unigraph (Hold), digraph (Down-Down, Down-Up, Up-Down, Up-Up), trigraph (Hold-Tri), and their combinations. Training and testing were conducted in the fifth and sixth phases using a Support Vector Machine (SVM) classifier, leveraging keystroke patterns for behavioral biometric identification. The final phase focused on evaluating the models. Each model was tested under two scenarios: one where only the first user is treated as the authorized user, and another where the first three users are considered authorized. Each scenario was further divided into two cases based on preprocessing conditions. The models were assessed using multiple performance metrics, including Accuracy, F1-Score, Recall, Precision, ROC-AUC, and Equal Error Rate (EER). The highest achieved results were Accuracy of 99.35%, F1-Score of 94.2%, Recall of 91.8%, Precision of 98.8%, ROC-AUC of 99.56%, and a minimum EER of 0.02. These outcomes demonstrate the effectiveness of the proposed approach in enhancing authentication reliability using keystroke dynamics.

groups

Rasha Khalid Ibrahim mail -

Mays M. Hoobi mail

link https://doi.org/10.54216/JISIoT.180106

Volume & Issue

Vol. Volume 18 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Deep Learning Techniques For Image Splicing Detection: A Systematic Review

Currently, images stand for a highly common form of communication, whether through teleconferencing, mobile communication or social media. The identification of counterfeit images is intrinsic because it is crucial that the images used for communication be genuine and original. Images are fabricated referring to the fact that it is challenging to set the difference between a tampered image and the real image. This refers notably to the myriad technological, moral, and judicial implications connected with advanced image editing software. The majority of handcrafted traits are used in traditional approaches for detecting image counterfeiting. The problem with many of the image tampering detection methods now in use resides in the fact that they are confined to identifying particular types of alteration by looking for particular features in the images. Image tampering is currently recognized through deep learning techniques. These methods proved to be promising and worthwhile as they perform better than traditional ones since they can extract complex components from images. As far as this research paper is concerned, we provide a thorough review of deep learning-based methods for detecting splicing images, along with the pertinent results of our survey in the form of findings and analysis.

groups

Mohammed S. Khazaal mail -

Mohamed Elleuch mail -

Monji kherallah mail -

Faiza Charfi mail

link https://doi.org/10.54216/JISIoT.180107

Volume & Issue

Vol. Volume 18 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Optimizing Neural Network Architectures with TensorFlow and Keras for Scalable Deep Learning

Deep studying architectures face fundamental demanding situations in balancing overall performance optimization, computational scalability, and operational interpretability. Current strategies show off an essential fragmentation: neural architecture search (NAS) techniques perform independently of interpretability requirements, while scalability answers remain detached from structure optimization pipelines. This disconnect hinders the improvement of a unified workflow from architecture layout to interpretable deployment. We endorse DeepOptiFrame, a TensorFlow/Keras-primarily based Python framework that combines three middle capabilities: (1) superior optimization algorithms (BOHB, Hyperband) with useful resource-restrained multi-objective search, (2) distributed training acceleration across GPU/GPU clusters via Horovod integration and blended-precision strategies, and (3) GPU-increased interpretability gear (SHAP, LIME) incorporated without delay into the education pipeline. Our framework demonstrates large experimental improvements: a 15-20% accuracy growth at the CIFAR-a hundred and ImageNet benchmarks compared to today's baselines, a 65% education speedup whilst scaled to eight GPUs with close to-linear performance, and a 30% development in interpretability reliability, as measured via the Mean Confidence Decrease metric. This implementation additionally reduces reminiscence intake via forty% throughout gradient checkpoints even as keeping numerical balance. These advances establish a new paradigm for coherent deep learning development, simultaneously improving overall performance, scalability, and transparency inside unified workflow surroundings.

groups

Muna Al-Saadi mail -

Bushra Al-Saadi mail -

Dheyauldeen Ahmed Farhan mail -

Oday Ali Hassen mail

link https://doi.org/10.54216/JISIoT.180108

Volume & Issue

Vol. Volume 18 / Iss. Issue 1

Details open_in_new