Jaskaran Singh

EDDSBS & VBSF-TLD: Advanced Drone Detection and Optimization Framework

First Author | IEEE Open Journal of the Communications Society
Part 1: EDDSBS - Ensemble-Based IoT-Enabled Drones Detection Scheme
Developed an ensemble-based drone detection model (EDDSBS) integrating advanced transfer learning techniques to enhance detection accuracy of IoT-enabled drones. By leveraging pre-trained models from related domains, the system significantly reduces the need for extensive training datasets. The architecture combines Convolutional Neural Networks (CNNs) and machine learning classifiers to ensure real-time precision in high-security zones like airports and military bases. Tested on prominent benchmark datasets such as Drone-vs-Bird and UAVDT, EDDSBS outperforms existing detection methods by incorporating IoT sensors with real-time analytics for improved security and privacy protection.

Part 2: VBSF-TLD - Validation-Based Soft Computing Approach for Drone Detection Building on EDDSBS, VBSF-TLD introduces a validation-based transfer learning framework that leverages soft computing principles. This sub-model incorporates Particle Swarm Optimization (PSO) with the FastRCNN detection framework, dynamically adjusting parameters during training to optimize localization and classification in complex environments. The system’s adaptability and accuracy were enhanced further through U-Net-based background subtraction, improving precision in challenging conditions. The dual approach of EDDSBS and VBSF-TLD ensures a comprehensive solution to drone detection, optimizing performance across various environmental conditions.
First Paper Preprint Second Paper Code

Computer Vision-Based Early Flood Prediction System for Rural Roads

Winner @ (Smart India Hackathon 2022) | Adopted by the Ministry of Rural Development, India
Designed and developed a cloud-hosted computer vision model integrated with an explainable AI module to predict early flood risks, specifically focused on rural roads and infrastructure. The system analyzed satellite imagery and extracted critical data, including road quality, climate conditions, forest cover, and cloud cover, to anticipate potential flood occurrences. A custom backend engine built using Python employed advanced mathematical models to derive elevation angles and drainage areas, further refining the model's capacity to predict flood-affected zones. The frontend was developed using Next.js, enhancing accessibility and scalability. This project won the Smart India Hackathon 2022 and was recognized by the Ministry of Rural Development, leading to its adoption as a disaster management tool to protect rural roads.
Code

TextAI 4.0 (Psych): Deep Learning Attention enabled Fused Multimodal paradigm for depression analaysis

First Author| Preprint
Developed "TextAI 4.0 (Psych)," a deep learning attention-enabled fused multimodal paradigm embedded with explainability for depression detection in cross-domain frameworks. The model integrates 28 AI sub-models, combining text-based transformers (BERT, ALBERT, GPT) and vision-based networks (ResNet, VGG, Inception) through a fused multimodal architecture. Leveraging self-attention mechanisms, the system focuses on relevant features across text and visual data, improving detection accuracy and cross-domain robustness with minimal training data. The model incorporates explainability tools like GradCam and LIME to ensure interpretability, demonstrating enhanced accuracy and F1 scores in real-time depression detection
Code

DermAI 1.0: A Robust, Generalized, and Novel Attention-Enabled Ensemble-Based Transfer Learning Paradigm for Multiclass Classification of Skin Lesion Images

Developer | MDPI (1st Version) | Preprint (2nd Version)
: Developed DermAI 1.0, an attention-enabled ensemble-based deep learning model for multiclass skin lesion classification. This system leverages seven pre-trained transfer learning (TL) models and combines advanced ensemble techniques: stacking, softmax voting, and weighted averages to improve diagnostic precision. Attention mechanisms were integrated into attention-enabled TL (aeTL) models, which were further optimized into attention-enabled ensemble-based deep learning (aeEBDL) models. The model incorporates explainable AI techniques Gradient-weighted Class Activation Mapping (GradCam), and Local Interpretable Model-agnostic Explanations (LIME), offering transparency into the decision-making process, enhancing trust and interpretability in clinical diagnoses. This framework delivers a robust and adaptive solution for skin lesion classification.
Paper Code

GeneAI 3.0: A Hybrid Deep Learning Framework for miRNA Species Classification with Advanced Feature Extraction

First Author| Nature Scientific Reports 2024
We developed GeneAI 3.0, an advanced deep learning framework designed for the classification of miRNA species, with a strong emphasis on feature extraction. The framework processes miRNA sequences, which consist of nucleotide bases A, C, U, and G, by analyzing their inherent co-occurrences and dependencies. To capture the intricate relationships between these nucleotides, we extracted five conventional texture features — Entropy, Dissimilarity, Energy, Homogeneity, and Contrast — from the nucleotide sequences. In addition to these conventional features, we integrated three contemporary features: Shannon Entropy, Hurst Exponent, and Fractal Dimension. Shannon Entropy quantifies the uncertainty and randomness within the sequence, while the Hurst Exponent evaluates long-term dependencies and self-similarity in the sequence. The Fractal Dimension was used to describe the complexity and irregularity of the sequence's structural patterns, enabling deeper insights into the miRNA's behavior. We transformed the nucleotide arrangements into high-dimensional vector combinations based on these extracted features. Each miRNA sequence was mapped into this feature space, allowing the model to leverage both conventional and contemporary characteristics for precise classification. Through leveraging ensemble deep learning (EDL) and machine learning (EML) techniques, GeneAI 3.0 demonstrated superior performance in both binary and multiclass classification tasks, outperforming traditional models in terms of accuracy and reliability. This model was benchmarked against several solo and ensemble models, showcasing its efficacy in gene sequence analysis, with extensive validation using explainable AI (XAI) to interpret the classification results.
Paper Code

Diabetic complication detection and mitigation using Knowledge base filtering

Joint Winner| Young Innovator's Award
We developed a comprehensive detection framework aimed at identifying diabetic complications such as diabetic retinopathy, kidney disease, and cardiovascular conditions. This framework utilized a VGG19-based convolutional neural network architecture in TensorFlow to enable precise pattern recognition from medical imaging and clinical data. The model was specifically fine-tuned to classify diabetic retinopathy from retinal fundus images and was expanded to include additional layers for predicting kidney dysfunction and heart conditions associated with chronic diabetes. To further enhance patient management, we implemented a knowledge-based filtering recommender system and uses Neo4j (Graph-DBMS) for constructing a semantic graph database. This system analyzes patient profiles, historical medical data, and treatment outcomes to recommend personalized interventions for mitigating the risks of diabetic complications. We also developed a React-based front-end application for patient interaction, providing users with real-time insights into their health metrics. The project was presented at the InnoHealth conference, where it was recognized with the Young Innovators Award for its innovative approach to detecting diabetic complications and delivering personalized healthcare solutions.
Code

Attention-Enabled Ensemble Deep Learning Models for Depression Detection

First Author| Diagnostics
In this project, we developed attention-enabled ensemble deep learning (aeEDL) models designed for depression detection, focusing on a cross-domain adoption framework. We integrated self-attention mechanisms within ensemble architectures to enhance the model’s ability to focus on critical input features, improving detection accuracy. The system utilizes a combination of multiple solo deep learning models such as BERT, BiLSTM, GRU, and CNN, merging their outputs for superior performance. The model was rigorously validated using both seen and unseen datasets across multiple domains, demonstrating an overall accuracy improvement of up to 5% when compared to solo deep learning counterparts. These advancements position our framework as a robust solution for identifying depression across diverse datasets in real-world applications.
Paper Code

ASCP-IoMT: AI-enabled Lightweight Secure Communication Protocol for the Internet of Medical Things (IoMT)

Collaboration with CoE-AIML and Sponsered by US Department of Defense| IEEE Access 2022
We developed ASCP-IoMT, an AI-enabled, lightweight, secure communication protocol designed for the Internet of Medical Things (IoMT) ecosystem. Our project focuses on enhancing the interaction between patients and healthcare providers by securely connecting devices and transmitting sensitive healthcare data over the internet. To address potential security concerns, we proposed an innovative scheme that integrates AI for predictive healthcare analytics, enabling the system to forecast phenomena such as the likelihood of heart attacks or the presence of tumors. We conducted both informal and formal security analyses, using the random oracle model, and demonstrated that our solution offers enhanced security and functionality over existing systems. Additionally, we implemented ASCP-IoMT to evaluate its impact on network performance, showing improvements in end-to-end delay and throughput across various scenarios. In the AI-based Big Data analytics phase, we applied decision trees, support vector machines, and logistic regression to healthcare data, demonstrating the effectiveness of AI-driven predictions in the IoMT environment.
Paper

SINN-RD: Spline interpolation-envisioned neural network-based ransomware detection scheme

First Author| Elsevier Computers and Electrical Engineering
Introduced SINN-RD, a cutting-edge ransomware detection scheme utilizing Spline Interpolation-envisioned Neural Networks to address the rising threat of ransomware. The project focuses on extracting valuable features from network metadata, including log files such as conn.log, ssl.log, and x509.log, which provide detailed insights into connection states, SSL certificates, and packet information. In this project, we proposed innovative mechanisms for data normalization and feature generation from these log files, enabling more precise ransomware detection. By leveraging metadata characteristics—such as certificate validity, cipher usage, and connection states—SINN-RD enhances the ability to distinguish between benign and malicious traffic with higher accuracy. These detailed extractions allow for a deeper analysis of ransomware behavior in real-time traffic. The security analysis demonstrated SINN-RD's robustness against multiple attack vectors, while practical implementation revealed improvements in accuracy, precision, and recall. Through advanced feature extraction from metadata, SINN-RD achieved superior performance compared to existing ransomware detection solutions, marking a significant advancement in ransomware mitigation strategies.
Paper Code