Data Reconstruction Attack on Federated Learning via Interpolation and Weighted Loss
A new data reconstruction attack targets horizontal Federated Averaging (FedAvg), the most widely used federated learning scenario. The attack uses an interpolation-based approximation to generate intermediate model updates from clients' local training, enabling reconstruction of private training data even when clients share parameters after multiple local steps. A layer-wise weighted loss function further improves reconstruction quality by assigning different weights to updates across layers. This work addresses a critical gap, as prior attacks mostly failed against FedAvg. The method is detailed in arXiv:2308.06822.
Key facts
- Federated Learning (FL) is a distributed learning paradigm for collaborative model building without sharing private data.
- Recent data reconstruction attacks can recover clients' training data from shared parameters.
- Most existing methods fail against horizontal Federated Averaging (FedAvg) where clients share parameters after multiple local training steps.
- The proposed attack uses an interpolation-based approximation to generate intermediate model updates.
- A layer-wise weighted loss function assigns different weights to model updates in different layers.
- The attack makes FedAvg scenarios feasible for data reconstruction.
- The research is published on arXiv with ID 2308.06822.
- The paper is a replace-cross announcement type.
Entities
Institutions
- arXiv