Privacy-Preserving Federated Deep-Equilibrium Learning for Medical Image Classification
Αλέξανδρος Γκίλλας, Dimitris Ampeliotis, Kostas Berberidis · 2024
In this work, we study the problem of skin cancer diagnosis from images by employing a network of collaborating institutions (e.g, hospitals) that cooperate under the emerging federated learning protocol. In such a scenario, the problems of not exposing sensitive patient information as well as the heterogeneity of the participating devices are of paramount importance. To this end, we propose the use of deep equilibrium models, in place of some other "traditional" deep neural network model, that offer a natural means of dealing with devices that have different computational resources. Furthermore, to prevent the leakage of sensitive information, the models exchanged in the proposed approach are homo-morpically encrypted. Numerical results indicate that the proposed approach offers the same accuracy as compared to the state-of-the-art federated learning case that "traditional" deep-learning models, but with three significant advantages: (a) increased privacy, (b) support of heterogeneous devices, and (c) significantly reduced communication requirements.