Digital innovation has emerged as a transformative force driving the growth and development of economies worldwide, particularly in developing nations. By leveraging technology, these countries can overcome structural challenges, improve productivity, and achieve sustainable economic progress. In an era defined by rapid technological advancements, digital tools and platforms are playing a pivotal role in fostering economic inclusion, enhancing public services, and creating opportunities for individuals and businesses alike. Developing economies often face hurdles such as limited infrastructure, inefficient systems, and high levels of poverty. Digital innovation offers solutions to these issues by introducing cost-effective, scalable technologies that address local challenges while fostering global competitiveness. Mobile connectivity, for instance, has revolutionized access to financial services in regions where traditional banking systems are underdeveloped.
Phishing attacks pose a major threat to cybersecurity by deceiving users into revealing sensitive information. This study proposes an optimized phishing detection framework that leverages both syntactic URL features and temporal behavior through a novel ensemble model combining Gradient Boosting and Recurrent Neural Networks (RNN). The Gradient Boosting component captures static patterns and feature importance from structured URL attributes, while the RNN processes sequential character-level URL data to learn underlying patterns of malicious domains. The ensemble approach enhances detection accuracy, reduces false positives, and adapts effectively to evolving phishing tactics. Experimental results on benchmark datasets demonstrate superior performance compared to individual models, confirming the potential of hybrid ensemble techniques for real-time phishing URL detection.
Physiological signal-based emotion recognition has emerged as a significant field of study with applications in interactive system interfaces, affective computing and healthcare. The DREAMER dataset to explore the effectiveness of EEGNet, a compact convolutional neural network designed for EEG-based interfaces between brain and computers in emotion classification. EEGNet can attain superior emotion recognition accuracy by learning directly from raw EEG data, thus outperforming traditional feature-based methods. The experimental results demonstrate that EEGNet attains accuracy for valence classification and arousal classification, significantly surpassing conventional machine learning approaches. This paper demonstrates the usefulness of EEGNet for creating EEG-based emotion identification systems as a versatile and effective solution for real-world application. The possible future concepts under this research include multimodal data integration and subject independent classification, which enhance deep learning in physiological signal processing.
This chapter provides a detailed overview of the security threats pertaining to federated learning, with a specific focus on label-flipping type of data poisoning attacks. Accordingly, the goal of this chapter is to evaluate the robustness of the defense mechanisms to defend against such label flipping attack. In this chapter, the defense approaches like manifold learning approach (UMAP), clustering method (K-Means), and dimensionality reduction methods such as PCA and KPCA are evaluated and their limitations are illustrated. A Multi-Stage Clustering and Data Quality Filtering (MSCDFL) approach is introduced in this chapter to address these limitations which integrates several techniques like adaptive clustering, data quality scored updates, delta-based updates, and trimmed mean aggregation to detect and mitigate the adversaries. From the experimental results, it is observed that MSCDFL exhibits better performance by maintaining robustness and significantly improving the prediction accuracy even in the presence of 40% of adversaries.
Underwater plastic pollution threatens marine ecosystems and biodiversity, requiring efficient detection methods. This study proposes a deep learning-based underwater plastic waste detection framework using YOLOv8 and compares its performance with YOLOv5. A labeled dataset containing underwater images of plastic waste, including bottles, bags, and wrappers, was used for training and evaluation. Experimental results show that YOLOv8 achieves higher precision, recall, mean Average Precision (mAP), and faster inference speed than YOLOv5. The proposed framework demonstrates its effectiveness for real-time underwater plastic waste detection and supports marine environmental monitoring, conservation, and sustainable ocean cleanup initiatives.