Privacy-enhanced federated learning via asynchronous aggregation and local differential perturbation
A study proposes a privacy-enhanced federated learning framework integrating Dynamic Differential Privacy, lightweight Homomorphic Encryption, and Local Differential Privacy. It employs asynchronous aggregation with version control. Experiments on CIFAR-10 and Purchase-100 show up to 82.6% accuracy at ε=0.1 and 21.3% lower communication overhead than FedAvg.
The research introduces a federated learning framework that combines multiple privacy mechanisms to enable secure collaborative training on distributed data. It uses asynchronous aggregation to handle non-synchronized updates and demonstrates strong privacy-utility trade-offs on standard benchmarks.
The combination of DDP, HE, and LDP with asynchronous aggregation suggests a layered defense strategy. The reported accuracy under strict privacy (ε=0.1) indicates effective noise calibration. The 21.3% communication reduction implies optimized update compression or scheduling.
This approach could lower barriers for cross-silo federated learning in privacy-sensitive sectors like healthcare and finance, where asynchronous participation is common. The reduced communication overhead may make federated learning more feasible on edge devices.
Enables collaborative model training without raw data sharing, potentially unlocking new data partnerships and reducing compliance risks. Efficiency gains may lower operational costs for distributed training.
Next signals include adoption in real-world federated learning platforms, further reductions in communication cost, and extensions to more complex models or non-IID data distributions.