Kuldeep Vayadande
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kuldeep Vayadande has a total of 165 co-authors. Here are the top co-authors by number of shared publications:
Yogesh Bodhe
Government Polytechnic
7 shared publications- Rahul Prakash Mirajkar
Bharati Vidyapeeth’s College of Engineering
4 shared publications - Yogesh Uttamrao Bodhe
Government Polytechnic, Pune, India
4 shared publications - Ajit R. Patil
Bharati Vidyapeeth's College of Engineering, Pune, India
3 shared publications Amol Bhosle
MIT Art, Design and Technology University
3 shared publications- Gitanjali Bhimrao Yadav
Vishwakarma Institute of Information Technology, India
3 shared publications
Latest Publications
- 29 July 2026
- Published by IGI Global Scientific Publishing
- p. 367-400
Real-Time Malware Detection Using System Call Sequence Classification
Contemporary malware is increasingly evading signature-based protection by modifying payloads and binaries; accurate detection therefore requires monitoring runtime activity rather than static artifacts. We introduce a low-overhead, hybrid detector with a Random Forest front end for quick, interpretable triage and a convolutional-LSTM sequence model for detecting temporal motifs in brief syscall windows. Both branches run in parallel on tight, per-process windows of user-space trace data; a small logistic meta-learner fuses their probability outputs into a single calibrated decision. The pipeline is separated; collection, feature construction, training, and inference are distinct. It is designed to keep per-decision latency and memory use low on standard Linux hosts. On hand-crafted syscall traces and stratified holdout tests, the system achieves balanced performance: overall accuracy is about 97.0%, precision is around 96.5%, recall is roughly 96.0%, and F1 is about 96.2%.
- 29 July 2026
- Published by IGI Global Scientific Publishing
- p. 367-400
Real-Time Malware Detection Using System Call Sequence Classification
Contemporary malware is increasingly evading signature-based protection by modifying payloads and binaries; accurate detection therefore requires monitoring runtime activity rather than static artifacts. We introduce a low-overhead, hybrid detector with a Random Forest front end for quick, interpretable triage and a convolutional-LSTM sequence model for detecting temporal motifs in brief syscall windows. Both branches run in parallel on tight, per-process windows of user-space trace data; a small logistic meta-learner fuses their probability outputs into a single calibrated decision. The pipeline is separated; collection, feature construction, training, and inference are distinct. It is designed to keep per-decision latency and memory use low on standard Linux hosts. On hand-crafted syscall traces and stratified holdout tests, the system achieves balanced performance: overall accuracy is about 97.0%, precision is around 96.5%, recall is roughly 96.0%, and F1 is about 96.2%.
- Open Access
- 1 July 2026
- Published by Institute of Advanced Engineering and Science in International Journal of Reconfigurable and Embedded Systems (IJRES)
- Vol. 15  (2)
Binary hybrid pathfinder algorithm for efficient feature selection in resource-constrained embedded systems
Feature selection is critical for embedded machine learning systems where computational resources and memory are severely constrained. This paper presents the binary quadratically interpolated hybrid pathfinder algorithm (BQIHPFA), a novel metaheuristic optimization method designed for efficient feature subset selection in resource-limited classification tasks. BQIHPFA adapts the continuous QIHPFA to binary search spaces through sigmoid transfer functions and employs a hybrid two-group enhancement strategy combining pathfinder dynamics with salp swarm algorithm-inspired exploration. We evaluate BQIHPFA against three established binary optimization algorithms (binary particle swarm optimization (BPSO), binary grey wolf optimizer (BGWO), and binary whale optimization (BWO)) on three benchmark datasets with varying dimensionalities: LÃngua Brasileira de Sinais (Brazilian Sign Language) movement (90 features), Parkinson's disease detection (22 features), and Sonar Rock vs. Mine (60 features). Experimental results demonstrate that BQIHPFA achieves competitive classification accuracy (average 83.57%) with substantial feature reduction (average 64.1%) while executing 5.2 times faster than complex baselines and consuming minimal memory (peak: 45-58 MB). Ablation experiments demonstrate that every algorithmic part makes a 8-24% contribution to the total performance. BQIHPFA offers an easy-to-use, non-specific feature selection method to automated resource-constrained embedded classification systems, applicable to be deployed to low-power computing environments, and internet of things (IoT) edge systems.
- 1 July 2026
- Published by Institute of Advanced Engineering and Science in International Journal of Reconfigurable and Embedded Systems (IJRES)
- Vol. 15  (2)
Open AccessBinary hybrid pathfinder algorithm for efficient feature selection in resource-constrained embedded systems
Feature selection is critical for embedded machine learning systems where computational resources and memory are severely constrained. This paper presents the binary quadratically interpolated hybrid pathfinder algorithm (BQIHPFA), a novel metaheuristic optimization method designed for efficient feature subset selection in resource-limited classification tasks. BQIHPFA adapts the continuous QIHPFA to binary search spaces through sigmoid transfer functions and employs a hybrid two-group enhancement strategy combining pathfinder dynamics with salp swarm algorithm-inspired exploration. We evaluate BQIHPFA against three established binary optimization algorithms (binary particle swarm optimization (BPSO), binary grey wolf optimizer (BGWO), and binary whale optimization (BWO)) on three benchmark datasets with varying dimensionalities: LÃngua Brasileira de Sinais (Brazilian Sign Language) movement (90 features), Parkinson's disease detection (22 features), and Sonar Rock vs. Mine (60 features). Experimental results demonstrate that BQIHPFA achieves competitive classification accuracy (average 83.57%) with substantial feature reduction (average 64.1%) while executing 5.2 times faster than complex baselines and consuming minimal memory (peak: 45-58 MB). Ablation experiments demonstrate that every algorithmic part makes a 8-24% contribution to the total performance. BQIHPFA offers an easy-to-use, non-specific feature selection method to automated resource-constrained embedded classification systems, applicable to be deployed to low-power computing environments, and internet of things (IoT) edge systems.
- Open Access
- 25 June 2026
- Published by Elsevier in SSRN Electronic Journal
Workload-Aware Vector Index Optimization for Approximate Nearest Neighbor Search in Multi-Vector Databases
ANN search serves as the core mechanism of current vector databases, however, its efficiency is largely affected by characteristics of data, required level of recall, memory constraints, and type of queries. Those issues raise further challenges when working with multi-vector databases where every single record consists of several vectors and queries might need to access various combinations of vector fields. To solve this issue, we propose MINT (Multi-vector Index Tuning), a framework for workload-aware ANN indexes optimization. MINT analyses workload patterns, generates candidates of indexes' configuration, calculates their estimated cost and recall based on sample-based modeling and selects an optimal plan with beam search optimization within specified recall and memory constraints. The proposed framework is developed in Python using libraries such as NumPy, hnswlib, scikit-learn, SciPy, and h5py, and delays complete HNSW indexes building until the optimal configuration is chosen. Experiments were conducted on synthetic six-column vector database, based on GloVe, SIFT, Deep1M, and Music100 datasets. Our experiments showed that the proposed approach provides the average recall improvement from 0.3438 to 0.9666 in comparison with a baseline per-column approach, while average query time increased from 2.92 ms to 8.41 ms.
- 25 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWorkload-Aware Vector Index Optimization for Approximate Nearest Neighbor Search in Multi-Vector Databases
ANN search serves as the core mechanism of current vector databases, however, its efficiency is largely affected by characteristics of data, required level of recall, memory constraints, and type of queries. Those issues raise further challenges when working with multi-vector databases where every single record consists of several vectors and queries might need to access various combinations of vector fields. To solve this issue, we propose MINT (Multi-vector Index Tuning), a framework for workload-aware ANN indexes optimization. MINT analyses workload patterns, generates candidates of indexes' configuration, calculates their estimated cost and recall based on sample-based modeling and selects an optimal plan with beam search optimization within specified recall and memory constraints. The proposed framework is developed in Python using libraries such as NumPy, hnswlib, scikit-learn, SciPy, and h5py, and delays complete HNSW indexes building until the optimal configuration is chosen. Experiments were conducted on synthetic six-column vector database, based on GloVe, SIFT, Deep1M, and Music100 datasets. Our experiments showed that the proposed approach provides the average recall improvement from 0.3438 to 0.9666 in comparison with a baseline per-column approach, while average query time increased from 2.92 ms to 8.41 ms.
- Open Access
- 3 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 95403-95427
NeuroBridge-AI: A Hybrid Biophysical and Machine Learning Framework for Multi-Scale Neural Simulation
Understanding how neural networks develop emergent behavior remains one of the main challenges in computational neuroscience. Models that balance the dynamics of microscopic ion-channels with macroscopic behavioral responses are important, but most strategies have lost cellular details in favor of network scale. OpenWorm c302 (302 neurons, ˜7000 synapses) and implementation of ModelDB (ModelDB 266796), a model of the human hippocampal circuit with healthy and epileptic variants, provide complementary advantages: connectome topology and complete neuronal physiology, respectively. This study offers a hybrid structure, which combines the hippocampal ModelDB-based biophysics with the network topology of c302, to allow real-time voltage propagation, synaptic co-ordination and temporal synchronization via Python-based cross-framework communication. Electro-physiological properties and locomotor emergent behaviors are reproduced in the hybrid simulation. The benchmark tests indicate a great correspondence with experimental results: AVA inter-neuron peak firing rate of 34 Hz (target 32-35 Hz), VB motor neuron synchrony index of 0.87 (target ≥ 0.85) and 96% correspondence in the sequences of body-wall muscle activation. The hybrid system maintains fine-grained neuronal fidelity and has realistic coordination on a network scale compared to stand-alone simulations of c302 or ModelDB. The condition of adaptation modulation of the synaptic weights and ion-channel conductance produces stable and non-saturating oscillatory dynamics. The hybrid system establishes a two way communication between the hippocampal and C. elegans neural models to study how biophysical and whole-network dynamics can interact to produce emergent behavior. The model combines a comprehensive model of neuronal dynamics with fully connected brain regions, which has led to research on how localized neurons processes can lead to complex organisms.
- 3 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 95403-95427
Open AccessNeuroBridge-AI: A Hybrid Biophysical and Machine Learning Framework for Multi-Scale Neural Simulation
Understanding how neural networks develop emergent behavior remains one of the main challenges in computational neuroscience. Models that balance the dynamics of microscopic ion-channels with macroscopic behavioral responses are important, but most strategies have lost cellular details in favor of network scale. OpenWorm c302 (302 neurons, ˜7000 synapses) and implementation of ModelDB (ModelDB 266796), a model of the human hippocampal circuit with healthy and epileptic variants, provide complementary advantages: connectome topology and complete neuronal physiology, respectively. This study offers a hybrid structure, which combines the hippocampal ModelDB-based biophysics with the network topology of c302, to allow real-time voltage propagation, synaptic co-ordination and temporal synchronization via Python-based cross-framework communication. Electro-physiological properties and locomotor emergent behaviors are reproduced in the hybrid simulation. The benchmark tests indicate a great correspondence with experimental results: AVA inter-neuron peak firing rate of 34 Hz (target 32-35 Hz), VB motor neuron synchrony index of 0.87 (target ≥ 0.85) and 96% correspondence in the sequences of body-wall muscle activation. The hybrid system maintains fine-grained neuronal fidelity and has realistic coordination on a network scale compared to stand-alone simulations of c302 or ModelDB. The condition of adaptation modulation of the synaptic weights and ion-channel conductance produces stable and non-saturating oscillatory dynamics. The hybrid system establishes a two way communication between the hippocampal and C. elegans neural models to study how biophysical and whole-network dynamics can interact to produce emergent behavior. The model combines a comprehensive model of neuronal dynamics with fully connected brain regions, which has led to research on how localized neurons processes can lead to complex organisms.
- 26 May 2026
- Published by Springer Nature
- p. 220-231
Predictive Modeling and Factor Analysis for Crop Yield Using SHAP-Based Explainable AI
- 26 May 2026
- Published by Springer Nature
- p. 220-231
Predictive Modeling and Factor Analysis for Crop Yield Using SHAP-Based Explainable AI
- 26 May 2026
- Published by Springer Nature
- p. 341-350
Robust Ensemble Meta-Learning for Superconducting Critical Temperature Prediction: Stacking Decision Trees and Boosted Models
- 26 May 2026
- Published by Springer Nature
- p. 341-350
Robust Ensemble Meta-Learning for Superconducting Critical Temperature Prediction: Stacking Decision Trees and Boosted Models
- Open Access
- 1 March 2026
- Published by Institute of Advanced Engineering and Science in International Journal of Reconfigurable and Embedded Systems (IJRES)
- Vol. 15  (1) , 224-235
Synaptic shield: fusion of ResNext–50 and long short-term memory for enhanaced deepfake detection
Recent developments in deepfakes have created much anxiety about the authenticity of any digital content and thus, calls for implementing detection mechanisms that will work accordingly. This paper uses Synaptic Shield, a innovative deep learning (DL) framework which is customized to detect alterations by deepfakes with high precision levels. It employs both convolution neural networks (CNNs) as well as modules for time feature extractions to test spatial and motion indicators from video data. High-level preprocessing pipelines in combination with confidence scoring mechanism help make Synaptic Shield adaptive toward manipulation techniques such as FaceSwap and DeepFake. The accuracy of our model surpasses other deepfake detection models with a high accuracy of 98.3%. The above results are based on exhaustive experimentation on standard datasets like FaceForensics++, DeepFake detection challenge (DFDC), and Celeb DeepFake (Celeb-DF). Synaptic Shield is shown to be the best with outstanding results that maintain a higher confidence score equivalent to its precision and reliability. Scalability in having the capacity to accommodate various manipulation techniques and levels of video quality indicates robustness in offering an effective method toward ensuring integrity in digital media. The work is an important move forward in addressing the problems created by DeepFake technologies.
- 1 March 2026
- Published by Institute of Advanced Engineering and Science in International Journal of Reconfigurable and Embedded Systems (IJRES)
- Vol. 15  (1) , 224-235
Open AccessSynaptic shield: fusion of ResNext–50 and long short-term memory for enhanaced deepfake detection
Recent developments in deepfakes have created much anxiety about the authenticity of any digital content and thus, calls for implementing detection mechanisms that will work accordingly. This paper uses Synaptic Shield, a innovative deep learning (DL) framework which is customized to detect alterations by deepfakes with high precision levels. It employs both convolution neural networks (CNNs) as well as modules for time feature extractions to test spatial and motion indicators from video data. High-level preprocessing pipelines in combination with confidence scoring mechanism help make Synaptic Shield adaptive toward manipulation techniques such as FaceSwap and DeepFake. The accuracy of our model surpasses other deepfake detection models with a high accuracy of 98.3%. The above results are based on exhaustive experimentation on standard datasets like FaceForensics++, DeepFake detection challenge (DFDC), and Celeb DeepFake (Celeb-DF). Synaptic Shield is shown to be the best with outstanding results that maintain a higher confidence score equivalent to its precision and reliability. Scalability in having the capacity to accommodate various manipulation techniques and levels of video quality indicates robustness in offering an effective method toward ensuring integrity in digital media. The work is an important move forward in addressing the problems created by DeepFake technologies.
- 19 February 2026
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 17  (3) , 138
Synergizing graph embeddings and metaheuristic optimization for fraudulent review detection
- 19 February 2026
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 17  (3) , 138
Synergizing graph embeddings and metaheuristic optimization for fraudulent review detection
- 20 January 2026
- Published by Springer Nature
- p. 355-367
A Generative AI and Machine Learning-Based Framework for Proactive CI/CD Pipeline Monitoring and Fault Remediation
- 20 January 2026
- Published by Springer Nature
- p. 355-367
A Generative AI and Machine Learning-Based Framework for Proactive CI/CD Pipeline Monitoring and Fault Remediation