Skin Disease Classification Using Privacy-Preserving Federated Learning.

Skin diseases, being one of the most common diseases worldwide, can occur to people of all ages and are caused by bacteria, infections, etc. Currently, skin diseases are initially diagnosed visually, which is often prone to errors. Skin diseases unable to be identified through inspection are identified using a biopsy process that uses dermoscopic analysis and is prescribed manually by physicians. However, a biopsy has its safety and accessibility issues, and manual inspection requires long periods of time. Therefore, this paper uses machine learning for image-based classification techniques for skin disease diagnosis. However, to be trained and tested, machine learning generally requires access to a dataset to be stored in a centralized server, which often raises many concerns regarding security and privacy. In a medical environment especially, maintaining the security and confidentiality of patients' records is very important. Therefore, with the increase in awareness of user privacy, this paper builds a federated learning system where data is decentralized. Using a dataset of more than 10,000 images, the federated learning system initially shows an overall accuracy rate of classifying skin diseases of about 79%. Since the original dataset has class imbalance problems, a data balancing technique is applied to enlarge the dataset and balance the samples per class in the dataset. After balancing the dataset, the performance of the classifier is improved significantly in that it achieves the classification accuracy of 95%. This system is shown to be effective for classifying the type of skin disease using image-based classification techniques, while also keeping usersensitive information secure. [ABSTRACT FROM AUTHOR]