Pratik Kanani
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Pratik Kanani has a total of 73 co-authors. Here are the top co-authors by number of shared publications:
- Mamta Padole
Faculty of Technology and Engineering, The Maharaja Sayajirao University of Baroda, Vadodara, India
6 shared publications Lakshmi Kurup
School of AI, Amrita University, Delhi, India
3 shared publicationsDeepali Patil
Dwarkadas J. Sanghvi College of Engineering, Mumbai, India
2 shared publicationsVedant Singh
2 shared publicationsVyom Unadkat
2 shared publications- Adit Kattukaran
Department of Artificial Intelligence (AI) and Data Science, Dwarkadas J. Sanghvi College of Engineering, Mumbai, India
1 shared publication
Latest Publications
- 1 April 2026
- Published by Springer Nature
- p. 141-155
VidGyan: AI-Powered Educational Video Generation from Text Prompts
- 1 April 2026
- Published by Springer Nature
- p. 141-155
VidGyan: AI-Powered Educational Video Generation from Text Prompts
- Open Access
- 25 October 2025
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 67-81
Evaluating and Improving Algorithms for Off-Road Routing
In this paper, we explore the multifaceted capabilities of route-finding algorithms and their role in delivering dynamic paths for diverse navigation scenarios. In off-road route finding, the shortest path is not always the best; the route's smoothness must also be considered. Present methodologies like A* have notable limitations, such as difficulty adapting to complex terrains and producing rugged routes. This limitation hampers performance in critical scenarios, such as in emergencies like landslides or earthquakes, where off-road exploration is crucial. This paper proposes 3 methodologies that have been fine-tuned with optimal hyperparameters for optimal results using gradient descent. Two of these methodologies, Simulated Annealing and Ant Colony Optimization, overcome some limitations of A* but not all. However, Q-Learning significantly overcomes the limitations and saves travel time by providing a route with 70% fewer undulations and a 16% reduction in route length compared to A*. Compared to other implementations, the Q-Learning implementation proposed in this research not only focuses on minimizing path length but also minimizes route undulations, providing a dual objective approach that is well suited to real-world scenarios. Unlike prior implementations, which focus on a single objective, such as path length or obstacle avoidance, this work leverages a reward function that penalizes elevation variance while rewarding shorter routes, resulting in smoother, more easily traversable paths. Thus, Q-Learning overcomes the cons of the present methodologies and can form synergistic combinations, enhancing the overall performance of off-road space searching systems and accommodating the various challenges and complexities inherent in the targeted applications. The dataset used includes features such as latitude, longitude, and elevation. The strategic application of heuristics enables the swift evaluation of multiple paths, facilitating the selection of optimal routes in real-time applications. By combining various heuristics, the development of off-road path identification systems capable of discerning optimal paths across varied terrains becomes feasible.
- 25 October 2025
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 67-81
Open AccessEvaluating and Improving Algorithms for Off-Road Routing
In this paper, we explore the multifaceted capabilities of route-finding algorithms and their role in delivering dynamic paths for diverse navigation scenarios. In off-road route finding, the shortest path is not always the best; the route's smoothness must also be considered. Present methodologies like A* have notable limitations, such as difficulty adapting to complex terrains and producing rugged routes. This limitation hampers performance in critical scenarios, such as in emergencies like landslides or earthquakes, where off-road exploration is crucial. This paper proposes 3 methodologies that have been fine-tuned with optimal hyperparameters for optimal results using gradient descent. Two of these methodologies, Simulated Annealing and Ant Colony Optimization, overcome some limitations of A* but not all. However, Q-Learning significantly overcomes the limitations and saves travel time by providing a route with 70% fewer undulations and a 16% reduction in route length compared to A*. Compared to other implementations, the Q-Learning implementation proposed in this research not only focuses on minimizing path length but also minimizes route undulations, providing a dual objective approach that is well suited to real-world scenarios. Unlike prior implementations, which focus on a single objective, such as path length or obstacle avoidance, this work leverages a reward function that penalizes elevation variance while rewarding shorter routes, resulting in smoother, more easily traversable paths. Thus, Q-Learning overcomes the cons of the present methodologies and can form synergistic combinations, enhancing the overall performance of off-road space searching systems and accommodating the various challenges and complexities inherent in the targeted applications. The dataset used includes features such as latitude, longitude, and elevation. The strategic application of heuristics enables the swift evaluation of multiple paths, facilitating the selection of optimal routes in real-time applications. By combining various heuristics, the development of off-road path identification systems capable of discerning optimal paths across varied terrains becomes feasible.
- Open Access
- 12 April 2025
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 27-41
Selective Hopping and Stable Path Integration to Boost AODV's Message Delivery Assurance
In addition to carrying communications, Vehicular Ad-hoc Networks can be utilized to transfer vital information across nodes in the network, potentially averting disastrous losses. By following the established standards, this vital information is transmitted by moving cars on the road in conjunction with parked cars. In this case, the communication is forwarded via a mediator known as a "roadside unit." The Ad-hoc On-Demand Distance Vector (AODV) routing protocol is one of the finest for such a wireless environment because it can tolerate variations in vehicle density and speed. Though it performs better than other comparable protocols, the AODV routing protocol does not show much promise when it comes to unreachable nodes, unstable forwarding paths, broadcasting storms, and short connection lifetimes. In order to improve efficiency, this research paper suggests adding multiple Road Side Units, modifying the conventional AODV routing protocol by adding selective hopping and the Delay Minimization Problem. To confirm the usefulness of the suggested model in terms of propagation delay, transmission loss, network lifetime, etc., it is also simulated. The findings gained support the superiority of the suggested strategy, laying the groundwork for its wider implementation.
- 12 April 2025
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 27-41
Open AccessSelective Hopping and Stable Path Integration to Boost AODV's Message Delivery Assurance
In addition to carrying communications, Vehicular Ad-hoc Networks can be utilized to transfer vital information across nodes in the network, potentially averting disastrous losses. By following the established standards, this vital information is transmitted by moving cars on the road in conjunction with parked cars. In this case, the communication is forwarded via a mediator known as a "roadside unit." The Ad-hoc On-Demand Distance Vector (AODV) routing protocol is one of the finest for such a wireless environment because it can tolerate variations in vehicle density and speed. Though it performs better than other comparable protocols, the AODV routing protocol does not show much promise when it comes to unreachable nodes, unstable forwarding paths, broadcasting storms, and short connection lifetimes. In order to improve efficiency, this research paper suggests adding multiple Road Side Units, modifying the conventional AODV routing protocol by adding selective hopping and the Delay Minimization Problem. To confirm the usefulness of the suggested model in terms of propagation delay, transmission loss, network lifetime, etc., it is also simulated. The findings gained support the superiority of the suggested strategy, laying the groundwork for its wider implementation.
- Open Access
- 25 February 2025
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 24-37
Comprehensive Analysis of Machine Learning and Deep Learning models on Prompt Injection Classification using Natural Language Processing techniques
This study addresses the prompt injection attack based vulnerability in large language models, which poses a significant security concern by allowing unauthorized commands by attackers to manipulate the outputs produced by model. Text classification methods used for detecting these malicious prompts are investigated on the prompt injection dataset obtained from Hugging Face datasets, utilizing a combination of natural language processing-based techniques applied on various machine learning and deep learning algorithms. Multiple vectorization approaches, like the Term Frequency-Inverse Document Frequency, Word2Vec, Bag of Words, and embeddings, are implemented to transform textual data into meaningful representations. The performance of several classifiers is assessed, on their ability to identify between malicious and non-malicious prompts. The Recurrent Neural Network model demonstrated high accuracy, achieving a detection rate of 94.74%. Obtained results indicated that deep learning architectures, particularly those that capture sequential dependencies, are highly effective in identifying prompt injection threats. This study contributes to the evolving field of AI security by addressing the issue of defending LLM based systems against adversarial threats in form of prompt injections. The findings highlight the importance of integrating sequential dependencies and contextual understanding in combatting LLM vulnerabilities. By the application of reliable detection mechanisms, this study enhances the security, integrity, and trustworthiness of AI-driven technologies, ensuring their safe use across diverse applications.
- 25 February 2025
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 24-37
Open AccessComprehensive Analysis of Machine Learning and Deep Learning models on Prompt Injection Classification using Natural Language Processing techniques
This study addresses the prompt injection attack based vulnerability in large language models, which poses a significant security concern by allowing unauthorized commands by attackers to manipulate the outputs produced by model. Text classification methods used for detecting these malicious prompts are investigated on the prompt injection dataset obtained from Hugging Face datasets, utilizing a combination of natural language processing-based techniques applied on various machine learning and deep learning algorithms. Multiple vectorization approaches, like the Term Frequency-Inverse Document Frequency, Word2Vec, Bag of Words, and embeddings, are implemented to transform textual data into meaningful representations. The performance of several classifiers is assessed, on their ability to identify between malicious and non-malicious prompts. The Recurrent Neural Network model demonstrated high accuracy, achieving a detection rate of 94.74%. Obtained results indicated that deep learning architectures, particularly those that capture sequential dependencies, are highly effective in identifying prompt injection threats. This study contributes to the evolving field of AI security by addressing the issue of defending LLM based systems against adversarial threats in form of prompt injections. The findings highlight the importance of integrating sequential dependencies and contextual understanding in combatting LLM vulnerabilities. By the application of reliable detection mechanisms, this study enhances the security, integrity, and trustworthiness of AI-driven technologies, ensuring their safe use across diverse applications.
- Open Access
- 30 October 2024
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 38-53
Enhancing Paediatric Pneumonia Detection and Classification Using Customized CNNs and Transfer Learning Based Ensemble Models
Pneumonia is one of the most prominent causes of mortality in children who are below the age of five years in most parts of the globe. Hence, adequate pneumonia diagnosis is of paramount importance and is what drove this research effort which has led to the development of two transfer learning-based ensemble models. One of the proposed models classifies the chest radiographs into normal and pneumonia cases with outputs being generated from VGG-16, Inception-v3, and two custom-made convolutional neural networks, PneumoNet-v1 and PneumoNet-v2. The second model distinguishes bacterial from viral pneumonia with the help of Xception, MobileNet-v2, and PneumoNet-v1. To accomplish the aim of the study, the Guangzhou Women and Children’s Medical Center dataset (Kermany Dataset) was used to benchmark model performance. PneumoNet-v1 and PneumoNet-v2 were designed with an emphasis for high classification accuracy and have individual accuracies of 96.2% and 96.8%, respectively for pneumonia detection. The first ensemble model used for classifying between healthy and infected images attained a classification accuracy of 98.03%. The second model used for differentiating between bacterial and viral demonstrated an accuracy of 91.93%. The effectiveness of transfer learning-based ensemble models as well as of the proposed custom CNN designs in enhancing the analysis of paediatric pneumonia and facilitating better diagnosis has been explored in this research.
- 30 October 2024
- Published by Asian Research Association in International Research Journal of Multidisciplinary Technovation
- p. 38-53
Open AccessEnhancing Paediatric Pneumonia Detection and Classification Using Customized CNNs and Transfer Learning Based Ensemble Models
Pneumonia is one of the most prominent causes of mortality in children who are below the age of five years in most parts of the globe. Hence, adequate pneumonia diagnosis is of paramount importance and is what drove this research effort which has led to the development of two transfer learning-based ensemble models. One of the proposed models classifies the chest radiographs into normal and pneumonia cases with outputs being generated from VGG-16, Inception-v3, and two custom-made convolutional neural networks, PneumoNet-v1 and PneumoNet-v2. The second model distinguishes bacterial from viral pneumonia with the help of Xception, MobileNet-v2, and PneumoNet-v1. To accomplish the aim of the study, the Guangzhou Women and Children’s Medical Center dataset (Kermany Dataset) was used to benchmark model performance. PneumoNet-v1 and PneumoNet-v2 were designed with an emphasis for high classification accuracy and have individual accuracies of 96.2% and 96.8%, respectively for pneumonia detection. The first ensemble model used for classifying between healthy and infected images attained a classification accuracy of 98.03%. The second model used for differentiating between bacterial and viral demonstrated an accuracy of 91.93%. The effectiveness of transfer learning-based ensemble models as well as of the proposed custom CNN designs in enhancing the analysis of paediatric pneumonia and facilitating better diagnosis has been explored in this research.
- 6 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
A Multimodal Approach to Multispeaker Summarization and Mind Mapping for Audio Data
With recent global events resulting in a shift in the work setting and way of life worldwide, most educational institutes and corporate organizations have been forced to conduct their day-to-day operations online. Lectures and meetings that are conducted online can be downloaded and users can store these recorded meetings. Owing to the latest technological advancements, storage capacities are no longer a hindrance when storing and using this data. However, storing these recorded meetings will only be helpful when users can access and browse them quickly. This paper proposes implementing a multi-speaker speech summarization model to benefit from the content captured in meeting recordings. Browsing through audio-visual data for information can be a time-consuming task for humans since manual summary generation would require a person to sit through the whole meeting. This can be overcome by combining Natural Language Processing techniques with text summarization to generate the transcript of a meeting. By performing speaker diarization, the various speakers in a multi-speaker meeting are identified, and accordingly, labels are assigned to every participant in the transcript. The model then performs summarization for individual speakers as well as the whole meeting and creates a mind map to represent the meeting minutes visually. Thus, this paper provides an efficient manner of reviewing meetings and labelling them for future reference.
- 6 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
A Multimodal Approach to Multispeaker Summarization and Mind Mapping for Audio Data
With recent global events resulting in a shift in the work setting and way of life worldwide, most educational institutes and corporate organizations have been forced to conduct their day-to-day operations online. Lectures and meetings that are conducted online can be downloaded and users can store these recorded meetings. Owing to the latest technological advancements, storage capacities are no longer a hindrance when storing and using this data. However, storing these recorded meetings will only be helpful when users can access and browse them quickly. This paper proposes implementing a multi-speaker speech summarization model to benefit from the content captured in meeting recordings. Browsing through audio-visual data for information can be a time-consuming task for humans since manual summary generation would require a person to sit through the whole meeting. This can be overcome by combining Natural Language Processing techniques with text summarization to generate the transcript of a meeting. By performing speaker diarization, the various speakers in a multi-speaker meeting are identified, and accordingly, labels are assigned to every participant in the transcript. The model then performs summarization for individual speakers as well as the whole meeting and creates a mind map to represent the meeting minutes visually. Thus, this paper provides an efficient manner of reviewing meetings and labelling them for future reference.
- 6 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 5, 1-6
Speech Recognition Email System for People with Impaired Sights
In today’s culture, where having access to the internet and digital communication is essential for daily life, it is crucial to recognize the importance of accessibility for those with visual disabilities. The proposed email system that utilizes speech recognition technology is a significant advancement that can greatly benefit individuals with impaired sight. It allows these individuals to easily compose, read, and navigate emails using audio and other non-visual means. The project’s contribution in enabling blind people to send and receive voice-based e-mail messages is commendable, and the system’s GUI has been evaluated and found to perform better than that of existing GUIs. The use of voice-to-text and text-to-voice techniques for accessing emails is also an important feature that makes it user-friendly for individuals with impaired sight. The fact that the system is based on interactive voice response makes it efficient and easy to use. The speech recognition email system represents a significant advancement for individuals with impaired sight, enabling them to stay connected with friends, family, and colleagues without the need for visual assistance. It is important to continue developing such technologies to ensure that digital communication is accessible to all individuals, regardless of their abilities.
- 6 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 5, 1-6
Speech Recognition Email System for People with Impaired Sights
In today’s culture, where having access to the internet and digital communication is essential for daily life, it is crucial to recognize the importance of accessibility for those with visual disabilities. The proposed email system that utilizes speech recognition technology is a significant advancement that can greatly benefit individuals with impaired sight. It allows these individuals to easily compose, read, and navigate emails using audio and other non-visual means. The project’s contribution in enabling blind people to send and receive voice-based e-mail messages is commendable, and the system’s GUI has been evaluated and found to perform better than that of existing GUIs. The use of voice-to-text and text-to-voice techniques for accessing emails is also an important feature that makes it user-friendly for individuals with impaired sight. The fact that the system is based on interactive voice response makes it efficient and easy to use. The speech recognition email system represents a significant advancement for individuals with impaired sight, enabling them to stay connected with friends, family, and colleagues without the need for visual assistance. It is important to continue developing such technologies to ensure that digital communication is accessible to all individuals, regardless of their abilities.
- Open Access
- 29 June 2023
- Published by Springer Nature in Research Square
PreciSplit: A Novel Approach to Predicting Polynomial Regression Data
Time complexity is crucial in machine learning models, especially with large datasets, as it impacts efficiency and scalability. Reducing time complexity even by a fraction is necessary to avoid computational bottlenecks and enable faster model development cycles. It saves on resources, allowing for cost-effective operations. Data splitting involves dividing datasets into subsets for training, validation, and testing. It enables unbiased evaluation, prevents overfitting, and facilitates hyperparameter tuning and model selection. By evaluating models on separate test sets, their performance on unseen data can be assessed, leading to better generalization and more reliable machine learning systems. This paper introduces PreciSplit, a highly effective approach for data splitting based on the dataset’s points of inflection. By strategically dividing the data at non-linear regions, PreciSplit creates nearly linear data segments, enabling the application of 1 precise linear regression on each split region. Additionally, the evaluation of models on distinct test sets allows for comprehensive assessment of their performance on unseen data. This evaluation process enhances generalization capabilities and fosters the development of more reliable machine learning systems. The proposed PreciSplit method showcases promising potential for advancing data analysis and model accuracy. For smaller datasets, PreciSplit shows slower training times and lower accuracy compared to Support Vector Regression (SVR).However, as the dataset size grows,it exhibits a substantial improvement in training speed, outperforming SVR, while maintaining only a slight increase in error margin.
- 29 June 2023
- Published by Springer Nature in Research Square
Open AccessPreciSplit: A Novel Approach to Predicting Polynomial Regression Data
Time complexity is crucial in machine learning models, especially with large datasets, as it impacts efficiency and scalability. Reducing time complexity even by a fraction is necessary to avoid computational bottlenecks and enable faster model development cycles. It saves on resources, allowing for cost-effective operations. Data splitting involves dividing datasets into subsets for training, validation, and testing. It enables unbiased evaluation, prevents overfitting, and facilitates hyperparameter tuning and model selection. By evaluating models on separate test sets, their performance on unseen data can be assessed, leading to better generalization and more reliable machine learning systems. This paper introduces PreciSplit, a highly effective approach for data splitting based on the dataset’s points of inflection. By strategically dividing the data at non-linear regions, PreciSplit creates nearly linear data segments, enabling the application of 1 precise linear regression on each split region. Additionally, the evaluation of models on distinct test sets allows for comprehensive assessment of their performance on unseen data. This evaluation process enhances generalization capabilities and fosters the development of more reliable machine learning systems. The proposed PreciSplit method showcases promising potential for advancing data analysis and model accuracy. For smaller datasets, PreciSplit shows slower training times and lower accuracy compared to Support Vector Regression (SVR).However, as the dataset size grows,it exhibits a substantial improvement in training speed, outperforming SVR, while maintaining only a slight increase in error margin.
- Retracted
- 14 June 2023
- Published by Springer Nature in Soft Computing
- Vol. 30  (S1) , 297
RETRACTED ARTICLE: Improving QoS of DSDV protocol to deliver a successful collision avoidance message in case of an emergency in VANET
- 14 June 2023
- Published by Springer Nature in Soft Computing
- Vol. 30  (S1) , 297
RetractedRETRACTED ARTICLE: Improving QoS of DSDV protocol to deliver a successful collision avoidance message in case of an emergency in VANET
- 24 March 2023
- Published by Springer Nature in SN Computer Science
- Vol. 4  (3) , 1-4
SQLIML: A Comprehensive Analysis for SQL Injection Detection Using Multiple Supervised and Unsupervised Learning Schemes
- 24 March 2023
- Published by Springer Nature in SN Computer Science
- Vol. 4  (3) , 1-4
SQLIML: A Comprehensive Analysis for SQL Injection Detection Using Multiple Supervised and Unsupervised Learning Schemes