Ph.D pursuing
Pillai’s College of Engineering, New Panvel (2021)
M.E. (Computer Engineering) – 72 %
Pillai’s College of Engineering, New Panvel (2013)
B.E. (Computer Engineering) – 68.60 %
St. Francis Institute of Technology, Borivali(2007)
Diploma (Computer Engineering) – 74.20 %
Sardar V Patel Polytechnic, Borivali (2003)
SSC – 81.15 %
High School, Virar(2000)
RESEARCH INTERESTS
Data base, Data Stream Mining, Artificial Intelligence
FUTURE PROJECTS
IoT Data stream Mining
Applications Invited
15
Scopus Publications
Scopus Publications
Color Space Fusion for Semantic Segmentation in Indian Adverse Weather Driving Scenes Veer Chheda, Vansh Shah, Hezal Lopes, Vishakha Shelke 6th Biennial International Conference on Nascent Technologies in Engineering Icnte 2026, 2026 Performing semantic segmentation in bad weather conditions is still extremely difficult because RGB sensors often have poor visibility in such conditions. In this study, we have explored various colour spaces to combine RGB and NearInfrared images to improve the perception of environment in bad weather conditions. We assessed the YUV, LAB and HSV colour spaces to combine RGB and NIR images along with an RGB baseline for comparison. The IDD-AW dataset was used for this fusion across the YOLOv8 and YOLOv11 nano and small segmentation models. According to our experimental results, LAB fusion along with YOLOv11 yields the best results which had a SmIoU of 17.96 and a mIoU of 26.46. These results are a 10% improvement over the RGB baseline. Since YUV and LAB color spaces have separate luminance channels, they have higher scores when compared to HSV, which shows moderate improvements. This tells us that choosing the right color for space for the RGB and NIR fusion has a huge positive impact on semantic segmentation in conditions with fog, low light, rain and snow even for smaller models focusing on real-time inference.
Benchmarking the Time Cost of Adversarial Attacks on GANs Khushi Sayani, Het Gohil, Hussain Degani, Hezal Lopes 6th Biennial International Conference on Nascent Technologies in Engineering Icnte 2026, 2026 Generative Adversarial Networks (GANs) are now a foundation of contemporary generative AI, driving everything from art to synthetic data generation. Yet, their susceptibility to adversarial attacks, as well as the rapidity with which these attacks can compromise them, is still a vital area of research. Although extensive research has been conducted on the success of attacks, there has not been a great deal of emphasis on the time it takes for an attack to be effective, a very important consideration in practical environments where prompt detection and prevention are absolutely essential. This paper introduces a benchmarking platform intended to quantify the time expense of GAN adversarial attacks. By considering three predominant attack categories, latent-space perturbations, training-time poisoning (backdoors), and membership inference, we not only consider their efficacy but also whether they can attack a model quickly. Our results indicate a speed vs. impact trade-off: the Latent Perturbation attack is quick to perform but infects individual outputs, whereas the time poisoning attack, take longer to seed but can infect models entirely or spill training data and finally Membership Inference attacks are quick to perform and can reveal if an individual’s data was used in a model’s training set. This benchmark seeks to offer a down-to-earth, time-savvy perspective for researchers and practitioners alike to more effectively evaluate GAN vulnerabilities and design timely defensive measures.
Real Time Drift Detection and Adaptation Using Hybrid ADWIN in Agricultural Environmental Monitoring System , Hezal Lopes, and International Journal of Electrical and Electronic Engineering and Telecommunications, 2025 Real-time analysis of streaming data is crucial in agricultural environmental monitoring to address quickly changing conditions like seasonal weather changes. Concept drift, where the statistical characteristics of input data evolve, poses a significant problem for static machine learning models. This research presents a drift-aware framework based on a hybrid adaptive windowing method combined with an Online Sequential Extreme Learning Machine (OS-ELM). The strategy involves a multidimensional extension of Adaptive Windowing (ADWIN) that is supplemented by the Kolmogorov–Smirnov statistical test and Hoeffding’s bound to identify and respond to realtime drift. An experimental Internet of Things (IoT) platform was constructed to gather environmental parameters such as temperature, humidity, soil moisture, light, pH, and rainfall. Empirical tests on real and synthetic datasets show that the new framework greatly enhances predictive performance, from 85.86 percent to 97.29 percent when drift handling is activated. The findings emphasize the significance of combining adaptive learning with drift detection for accurate and dependable prediction in precision agriculture.
Modified Deep ELM to Detect and Adapt Concept Drift in Data Stream Hezal Lopes, Prashant Nitnaware 2025 5th International Conference on Intelligent Technologies Conit 2025, 2025 Extreme learning machine (ELM) is a feedforward neural network (SLFN) which works with a single hidden layer. ELM is popularly used in data stream classification due to its speed and accuracy. In this paper, a modified deep extreme learning machine (ELM) is presented. In this proposed Deep ELM classifier, a concept drift detection method is integrated to detect change in data pattern in data stream. The experimental results showed the modified ELM algorithm improves the accuracy of classification as well as can adapt to new concepts in a very short period of time. Experiments were carried out with Agarwal data set. Comparative analysis between Extreme learning machine (ELM), Online Sequential ELM (OS-ELM) and the Modified deep ELM shows that Modified deep ELM significantly improves the classification accuracy, robustness and stability and offer reliable solutions for real time data stream applications.
Deepfake Detection Using Mesonet and EfficientnetB4 Fusion Model Shubham Bhimani, Srushti Borvadkar, Ayush Agrawal, Hezal Lopes 2025 International Conference on Applications of Machine Intelligence and Data Analytics Icamida 2025, 2025 There have been serious concerns about the advent of deepfake technology that employs sophisticated machine learning processes to produce hyper-realistic doctored media. Deepfakes have the potential to make disinformation easier to spread and create serious cybersecurity threats. While current deepfake detection processes are beneficial in certain contexts, there are certain limitations such as poor generalization to other datasets and vulnerability to adversarial attacks. In response to these concerns, we propose a new hybrid deepfake detection model that leverages the strengths of MesoNet and EfficientNet. MesoNet, extremely popular for its ability to extract handcrafted features, excels in detecting fine-grained evidence of media forgery by identifying convolutional hierarchies reflecting subtle inconsistencies. EfficientNet, a toprated deep network, is better at the same energy efficiency, making the model's output stable with no high computational cost. By combining these two sound models, our hybrid model aims to improve deepfake detection accuracy, overcome current limitations, and offer a sound solution to real world deployment where detection needs to be efficient and effective.
Portfolio Optimization Using Transformer Model Krishit Mehta, Priyansh Kothari, Sacchit Wathe, Parv Karia, Hezal Lopes 2025 International Conference on Applications of Machine Intelligence and Data Analytics Icamida 2025, 2025 The important factor is accurately predicting stock prices; accuracy is proven difficult due to the volatility of the market itself, non-linear functioning dependencies, and the dynamic changing of economic parameters. Traditional models, be they statistical regression or recurrent neural networks, have great difficulties accounted with capturing the complex patterns of the market and also in adapting very rapidly to price changes. To remedy this apprehension, the present research proposes a Transformer-based deep learning framework that to improve the accuracy of the prediction and carry out portfolio optimization uses self-attention mechanism and multi-dimensional feature representation. The model examines historical stock prices at a rolling window of sixty days, normalizes the data using MinMaxScaler, and identifies key features through a multi-head attention mechanism with position embeddings. The training is performed with Adam optimizer and Mean Squared Error (MSE) loss optimized for 50 epochs with early stopping and learning rate modification. The performance of the model is evaluated and compared using accuracy, precision, recall, F1-score, and ROC - AUC score across different stock datasets. Experimental results show prediction accuracy at impressive levels: 87.63 % for FB (Meta), 85.92 % for DAL (Delta Airlines), and 83.56 % for DELL. Confident buy/sell signals as high 80 % for high volume stocks were produced with the model, hence the dependence of such a result on a particular real-world trading application. In addition, risk-aware portfolio optimization is included with advantages of Monte Carlo simulations while using Sharpe Ratio analysis for dynamic asset allocation and adaptability to the market changes. Finally, merging predictive modelling with realtime risk management delivers a scalable, AI-driven investment approach to improved financial decision-making and risk reduced in volatile environments for maximized profitability.
Mind Matters: AI Driven Advanced Diagnostic of Mental Health Shrey Gandhi, Vidhi Sheth, Shruti Sahu, Rishabh Singh, Hezal Lopes 2025 IEEE 2nd International Conference on Green Industrial Electronics and Sustainable Technologies Giest 2025, 2025 Mental health disorders are becoming a global concern and there is a need for accessible, reliable and secure diagnostic tools to encourage early detection and intervention. Traditional mental health assessments rely on in-person evaluations and self-report surveys which may be time consuming, cost prohibitive and sometimes biased. Existing digital solutions often lack advanced security mechanisms and fail to provide personalized therapeutic guidance. The research outlines the development of a comprehensive medical diagnostic platform that integrates machine learning algorithms and principles of user-centric design. Founded in 2012 the proposed platform evaluates user responses to a set of well-designed psychological questions, calculates scores relevant to different conditions and rates the severity of disorders such as ADHD, OCD, clinical depression, PTSD, dementia, schizophrenia, bipolar disorder, and anxiety disorders. It also offers an intuitive GUI that allows easy interaction, provide real-time feedback a large array of colors and graphic representations of result. In addition, it provides individualized therapeutic suggestions and connects users with qualified mental health professionals. It utilizes technologies like random forests, support vector machines (SVM), natural language processing techniques to ensure accurate diagnostics, data privacy, and improved user trust. A Linear Regression model demonstrated high accuracy in detecting mental health disorders while an SVM model performed well in severity classification. Preliminary testing on the effectiveness, reliability, and accessibility of the tool proves very promising in dealing with mental issues and proactive care.
Decentralized Ticketing: A Blockchain Solution for Secure and Fair Event Access Janhavi Nate, Mahek Parmar, Parth Shethia, Preem Pareva, Hezal Lopes 2025 International Conference on Sustainable Technologies for Humanity and Smart World Hswtech 2025, 2025 This research presents a blockchain market for local networks that resolves inefficiencies in traditional marketplaces based on third-party intermediaries for operations. Current systems are more likely to result in higher costs, low transparency levels and less control over transactions by users. The proposed system utilizes blockchain technology to offer a secure and efficient market for the purchase and sale of assets. Smart contracts offer automatic transactions, wherein agreements are executed in the absence of human beings. The platform offers a user-friendly interface to list, buy and sell assets. Payment processing using a blockchain makes information consistent and more resistant to fraud or manipulation. Results show that such a method supports the processing of transactions and usability over traditional systems. The site also has artist and creator tools available to monetize digital assets with tokenized ownership. Future extension might include integration into virtual worlds and broader asset classes to facilitate trading in more general usage scenarios. This document presents the need for blockchain-based marketplaces, reviews current work, identifies gaps and how the proposed model increases efficiency and accessibility within the local trading regimes.
File Sharing System using Cryptographic Hash Raj Shah, Gunj Waghela, Yashvi Soni, Priyanshi Jain, Hezal Lopes Proceedings of 3rd International Conference on Intelligent Cyber Physical Systems and Internet of Things Icoici 2025, 2025 This research proposes a decentralized file-sharing system that seeks to overcome the major security and privacy limitations of traditional centralized systems. By integrating blockchain, Inter Planetary File System (IPFS), and encryption techniques, the intended system facilitates confidentiality, integrity, and authenticity of information. Easily hackable centralized servers with corruptibility and tampering are the foundation of standard file-sharing systems. They further suffer from bad scalability, heavy maintenance, and substandard authentication mechanisms that increase the risk of credential theft. To address these issues, our framework introduces Bcrypt hashing as a form of password storage with adaptive computational complexity, whose brute-force attacks become much more difficult to perform in the long run. Blockchain-based authentication also protects tamper-proof and transparent user verification, which keeps credentials secure against credential-stuffing and similar threats. The key contributions of this work are: A tamper-resistant, privacy-ensuring architecture for file sharing. Enhanced authentication security with combination of Bcrypt and blockchain. Improved scalability and storage efficiency through integration with IPFS. The proposed model is better than traditional and existing decentralized systems by offering an adjustable, secure, and user-centric mechanism for secure file sharing. It shows that decentralized storage, when coupled with strong authentication and access control, can be a feasible and improved alternative to legacy systems.
IOT Based Weather Intelligence Jigar Parmar, Trishal Nagda, Pranay Palav, Hezal Lopes 2018 International Conference on Smart City and Emerging Technology Icscet 2018, 2018
Application H-secure for mobile security Hezal Lopes, Madhumita Chatterjee 2014 International Conference on Circuits Systems Communication and Information Technology Applications Cscita 2014, 2014
