Protecting patient data is critical in the healthcare sector. With the growth of digital applications and the increasing volume of patient data, there's a strong need for models that responsibly handle data while providing accurate results. This research investigates distributed machine learning techniques like Federated Learning (FL) and Split Learning (SL). Applying these to the Breast Cancer Wisconsin (Diagnostic) dataset, we achieved 98% accuracy with FL and 99% with SL, surpassing traditional centralized approaches. FL and SL deliver high accuracy while keeping data decentralized, enhancing patient data security. This research highlights the potential of these approaches in modern healthcare, where balancing data confidentiality and diagnostic precision is essential.

Read more View Publication Other Documents

Advancements in medical imaging have been significantly driven by deep learning technologies, particularly Convolutional Neural Networks (CNNs). However, dataset imbalances pose a challenge, as underrepresented medical conditions can lead to biases in diagnostic models. This research tackles this issue by leveraging Generative Adversarial Networks (GANs) for data augmentation in chest radiography. Using the ChestXray2017 dataset, which is heavily skewed towards pneumonia cases, Deep Convolution Generative Adversarial Networks (DCGAN) were employed to generate synthetic images of normal chest X-rays, balancing the dataset. The performance of a CNN was then compared before and after augmenting the data with these synthetic images. Initially, the CNN achieved 93% training accuracy and 87% validation accuracy on the un-augmented dataset. After adding 400 synthetic normal chest X-rays, training accuracy rose to 95%, while validation accuracy improved to 89%, indicating better generalization due to a more balanced dataset. Although GAN-based augmentation appears effective for addressing class imbalances, further research is needed to understand the quality and ethical implications of using synthetic images. Overall, integrating GAN-generated images into CNN training provides a promising method to improve classification performance in medical imaging, offering a practical solution to data scarcity and imbalance.

Read more View Publication Other Documents

In our current era where data privacy is crucial, particularly in machine learning, our research presents a novel method that combines federated learning's privacy features with transformer-based models, designed specifically for recommendation systems. Federated learning offers a decentralized method that improves user privacy and data security. We utilize two transformer models, BERT and BST, within this framework. Analyzing their performance on the Amazon Customer Review dataset, the results are impressive: federated BERT achieves 80% accuracy, outperforming the standard model’s 65%, while federated BST reaches 94% accuracy, surpassing its traditional version's 82%. This study not only demonstrates federated learning’s ability to enhance accuracy but also its importance in protecting user privacy, marking a significant advancement in ethical AI.

Read more View Publication Other Documents

Comprehending and forecasting air quality is critical for public health and environmental governance, particularly in urban settings such as Delhi. Our research employs an extensive dataset from the Central Pollution Control Board, covering air pollutant levels at the ITO, Delhi station from 2017 to 2023. We concentrate on pollutants like PM2.5, PM10, NO2, NH3, SO2, Ozone, and CO, and address the prevalent issue of data gaps in environmental research. We tested various imputation methods, including KNN, Linear Regression, and Forward Fill, and found Forward Fill to be most effective for Ozone. Using this refined dataset, we then applied several models to forecast PM2.5 levels, employing methods like LSTM, ARIMA, and Bi-LSTM. The evaluation of these models, using daily pollution data, revealed that Bi-LSTM and LSTM with Attention were the most accurate in forecasting PM2.5 levels. The outcomes of this study are significant, providing valuable insights for environmental policies and health advisories in Delhi. By validating specific imputation and forecasting methods, our research aids in enhancing air quality monitoring and predictions. The success of the Bi-LSTM and LSTM with Attention models points to promising paths for future research focused on improving the accuracy of environmental forecasts.

Read more View Publication Other Documents

This research outlines the development, deployment, and evaluation of a sensor-based air quality monitoring system designed for kitchen settings. The system incorporates various sensors, including CCS 811, GP2Y1010, MQ135, MQ7, and DHT11, to monitor pollutants like carbon monoxide (CO), nitrogen oxides (NOx), particulate matter (PM), as well as temperature, humidity, total volatile organic compounds (TVOCs), and equivalent carbon dioxide (eCO2). Data gathering was conducted using the ThingSpeak platform, capturing readings across a 24-hour cycle in typical indoor spaces and a well-ventilated hostel mess kitchen with exhaust systems and air blowers. Despite advanced ventilation, findings revealed significantly elevated levels of TVOCs, eCO2, and dust in the kitchen compared to other indoor areas. Specifically, eCO2 and TVOC levels in the kitchen were up to two and ten times higher, respectively, than those indoors. This indicates that common exhaust systems are insufficient for removing the high volumes of VOCs produced during cooking. The study not only validates the monitoring system's effectiveness in tracking pollutants in real time but also points to the urgent need for better ventilation strategies in kitchens to improve air quality.

Read more View Publication Other Documents