Distributed Machine Learning on Differentially Private Skin Cancer Data
Machine Learning tasks are often hungry for vast amounts of data to be able to maximise performance. Utilising large amounts of real-world health data requires addressing issues concerning privacy. To address this problem, Google brought about advancements in “Federated Learning”  - a distributed approach for Machine Learning, which allows model training locally across multiple devices or nodes. Updated models are collected from these devices and aggregated to generate a global model. Notably, this approach requires data to be decentralized, hence, advocating privacy. Such an approach introduces a variety of challenges on the server side, the model inversion attack in particular, which allows image reconstruction from the received model weights. In this paper, we implement a FL (Federated Learning) setup for Skin Cancer classification and analyse the performance of various models by applying Differential privacy, which has proven to be able to defend against such types of attacks.