Blockchain-based Efficient Federated Learning Framework for Pneumonia Detection

Abstract

Blockchain-based Efficient Federated Learning Framework for Pneumonia Detection Abstract: Scientific advancements, such as IoT, Blockchain technology, and Artificial Intelli- gence, are reshaping the healthcare landscape, leading to the development of intelligent healthcare infrastructure. This progress aims to enhance the quality of medical care and extend the life expectancy of individuals. Medical professionals leverage AI techniques to diagnose and treat various critical conditions. However, ensuring patient privacy while harnessing the power of data-driven solutions poses a challenge due to strict privacy regulations in the healthcare sector. Pneumonia, a potentially life-threatening lung infection, stands as one of the primary concerns in healthcare. It occurs when the lungs become infected and inflamed, leading to difficulty breathing as they cannot function properly. Common symptoms include coughing, fever, and fatigue. To address the challenges of diagnosing pneumonia while safeguarding patient privacy, we explore the potential of decentralized training with federated learning combined with differential privacy. FL offers a decentralized approach to training models, allowing medical institutions to collaborate without sharing sensitive patient data directly. Blockchain technology plays a crucial role in managing the decentralized federated learning process. Leveraging its properties of transparency, immutability, and secure data storage, blockchain ensures the integrity and security of the FL-trained models. Through smart contracts, FL-trained weights are securely stored, addressing concerns related to data tampering. For the detection of pneumonia, we employ FedAvg, a pre-trained model modified for federated learning. Additionally, the blockchain component enables users to trace the decision-making process, providing transparency regarding the models and datasets utilized. In this study, we implemented decentralized training through federated learning to analyze various chest X-ray features for pneumonia prediction. Utilizing an image dataset containing both normal and pneumonia chest images, we distributed it across three dis- tinct FL clients for classification purposes. Our approach involved the utilization of a convolutional neural network that was trained and optimized in a decentralized manner to enhance pneumonia detection iteratively. We present the training accuracy achieved. here, which is 94.2%, showcasing the effectiveness of the distributed training approach. To safeguard the trained FL weights against data tampering, we incorporated key aspects of blockchain technology, including transparency, immutability, and secure data storage. Through a smart contract, the FL-trained weights are securely stored, addressing concerns regarding data integrity.