Beyond LDP-FL: Independent Uplink-Downlink Privacy Protection in Federated Learning

Federated learning (FL) has emerged as a key distributed intelligence paradigm in modern computer networks, enabling model training across massive edge devices without exposing raw data. Nevertheless, the continuous sharing of local model updates in FL may result in privacy exposures. To address this, Local Differential Privacy Federated Learning (LDP-FL), which injects noise on each client, was proposed. However, noise accumulates through LDP's two-phase indivisible sequential process, bringing the well-recognized privacy-utility trade-off. In this paper, to investigate whether LDP achieves a favorable trade-off, we propose an ideal interaction mode, Ideal Differential Privacy Federated Learning (IDP-FL), allowing independent protection in both uplink and downlink phases. Through comparative analysis, we discover and theoretically prove that LDP-FL suffers from inherent noise redundancy, i.e. noise accumulation in uplink exceeds privacy requirements in downlink. To avoid this, we propose Noise Annihilation Differential Privacy Federated Learning (NADP-FL) as an instantiation of IDP-FL. In this framework, paired noises are distributedly generated, mutually canceling during aggregation and not appearing in the downlink. Consequently, NADP realizes independent phase protection, eliminating unnecessary noise accumulation and achieving a more favorable privacy-utility trade-off without further utility loss. We validate its superior utility, scalability and robustness through extensive experiments.