FedCVESA: Taking Away Training Data in Federated Learning via Correlation Value Encoding and Segmented Aggregation

This paper studies a white-box Taking Away Training Data (TATD) attack against federated learning in which a malicious server actively writes client data into the global model. FedCVESA extends Correlation Value Encoding Attack (CVEA) with a Pearson-correlation regularizer applied at selected target clients, encoding private examples into dispersed carrier parameters. It then uses segmented aggregation to preserve those parameters while averaging the rest normally. Controlled experiments on MNIST, Fashion-MNIST, and CIFAR-10 with Dirichlet non-IID partitions reportedly recover semantically meaningful private images while retaining acceptable main-task utility.