Title: Backbone-as-Client Federated Learning with Differential Privacy for Diabetic Foot Ulcer Image Classification
Description:
This project develops a privacy-preserving federated learning framework for diabetic foot ulcer image classification. Instead of sharing raw medical images, pretrained CNN backbones such as VGG19, ResNet50, and DenseNet121 are used as local feature extractors, and only model updates are shared during federated training. The framework combines FedAvg, differential privacy, and protected update transmission to improve data privacy in medical image analysis. The study focuses on binary DFU classification and evaluates multiple MLP-based classification heads under non-IID client settings. This work highlights the potential of federated learning for privacy-aware diabetic foot ulcer screening while acknowledging the need for larger clinical datasets and external validation.