Generalizable Intrusion Detection System using Multimodal Machine Learning

• Designed a multimodal intrusion detection system that integrates Convolutional Neural Networks (CNNs) and Artificial Neural Networks (ANNs) to handle diverse data inputs from multiple dimensions and modalities.
• Developed a feature fusion mechanism that converts heterogeneous data sources into a unified feature space, enabling a single ML model to make accurate predictions on multimodal data.
• Optimized the model for lightweight deployment on edge devices, ensuring low latency, high efficiency, and real-time threat detection capabilities.
• Implemented adaptive anomaly classification algorithms, improving detection accuracy for cybersecurity threats across complex network environments.
• Validated the system on real-world datasets, achieving high performance in detecting cyber intrusions with minimal false positives.

End-to-End Explainable Intrusion Detection System with Self-Healing Capability Using Large Language Models (LLMs), XAI, and Machine Learning

• Developed a self-healing Intrusion Detection System capable of detecting cyber threats, identifying failure cases, and autonomously improving itself.
• Designed an extensive real-world data processing pipeline that feeds structured and unstructured security data into a robust ML model.
• Integrated Explainable AI (XAI) techniques to analyze incorrect classifications and identify failure points in model decision-making.
• Leveraged Large Language Models (LLMs) to provide insights into misclassified cases, enabling an automated retraining mechanism.
• Implemented a continuous learning framework where the IDS retrains itself periodically when a sufficient test dataset is accumulated.
• Utilized A/B testing to compare multiple trained models and dynamically select the most effective one for deployment.

Multidimensional Feature Learning Enhancement in IoT Intrusion Detection

• Developed a novel feature learning technique utilizing different layers of an autoencoder to train machine learning models effectively.
• Implemented a weighted averaging technique to aggregate classification results, improving the overall accuracy of IoT intrusion detection.
• Achieved high detection performance with enhanced feature extraction, leading to better generalization across datasets.
• Demonstrated that the learned feature representations can be transferred to different datasets for improved adaptability in IoT security applications.
• Published in the IEEE 10th World Forum on Internet of Things (WF-IoT).

Read the full paper: IEEE Xplore Link

UAV-Based Efficient and Reliable Wildfire Detection on the Edge

• Designed an autonomous wildfire detection system using Unmanned Aerial Vehicles (UAVs) equipped with advanced computer vision technology.
• Utilized the Flame2 dataset, which integrates RGB and infrared (IR) multispectral imagery to enhance fire detection under various environmental conditions.
• Developed a modified YOLOv8 (You Only Look Once) convolutional neural network to accurately detect wildfires in challenging scenarios such as smoke, nighttime, and dense vegetation.
• Achieved state-of-the-art accuracy with a 99.58% wildfire classification accuracy and an F1-score of 99.76%, significantly outperforming existing wildfire detection systems.
• Optimized the model for deployment on edge devices, ensuring real-time wildfire detection with minimal latency and high reliability.
• Published in the IEEE 10th World Forum on Internet of Things (WF-IoT).

Read the full paper: IEEE Xplore Link

Fraud Monitoring in Banking Systems Using Splunk & Red Hat Automation

• Conducted in-depth research on Splunk's real-time analytics for fraud detection in banking, focusing on log analysis, anomaly detection, and security event correlation.
• Explored use cases of Splunk in fraud monitoring, including transaction pattern analysis, risk scoring, and threat intelligence integration to enhance security measures.
• Analyzed the role of Red Hat automation in streamlining compliance processes, improving system resilience, and optimizing security workflows in financial institutions.
• Investigated how automation and SIEM tools can enhance fraud detection efficiency, reducing manual investigation time while improving fraud prevention accuracy.

AI and Data Science Research for Bangladesh’s First E-Healthcare System

• Worked as an Undergraduate Research Assistant focusing on AI and Data Science to develop Bangladesh’s first e-healthcare system.
• Researched and implemented machine learning models for COVID-19 diagnosis, knee osteoarthritis detection, and autism spectrum disorder (ASD) prediction.
• Developed models for Human Activity Recognition (HAR) using ensemble learning techniques to improve classification accuracy in healthcare applications.
• Designed a multimodal approach integrating image and tabular data for enhanced medical predictions.
• Optimized deep learning models to improve diagnostic accuracy while ensuring computational efficiency for deployment.
• Worked on data preprocessing, feature engineering, and model evaluation to ensure high generalizability across diverse patient datasets.
• Collaborated with healthcare professionals to fine-tune models based on real-world clinical data and feedback.
• Explored ensemble machine learning techniques to enhance classification accuracy in medical diagnostics.
• Applied statistical techniques to analyze model outputs and ensure interpretability in a clinical setting.

Google Scholar: View Research