This paper presents a novel data reconstruction attack method to address the vulnerability of a malicious central server in Federated Learning (FL) where it can reconstruct a client's private data. We overcome the limitations of existing methods, which rely on assumptions about the client data distribution and have poor efficiency at small batch sizes. By leveraging a novel geometric perspective on fully connected layers, we generate malicious model parameters that can perfectly reconstruct data batches of arbitrary size without prior knowledge of the client data. Experiments on image and tabular datasets demonstrate that our method outperforms existing methods, achieving perfect reconstruction of data batches two orders of magnitude larger than the previous best-performing method.
