FoggyTrust boosts federated learning robustness by 50% on CIFAR-10
New hierarchical trust model beats attacks in distributed ML without accessing private data.
Byzantine-robust federated learning aims to protect distributed model training from malicious clients without exposing private data. Existing methods like FLTrust use a trusted server-side root dataset to assign trust scores for aggregation, but struggle with globally heterogeneous data distributions. Researchers at arXiv propose FoggyTrust, a hierarchical extension that localizes trust computation to fog nodes—intermediate computing layers between clients and the server. This allows the framework to handle non-IID data by preserving robustness within locally homogeneous client groups while addressing distribution mismatch and client drift across groups using heterogeneity-aware optimizers such as FedAdam and SCAFFOLD.
In benchmark tests, FoggyTrust achieved its strongest gains on challenging heterogeneous settings. On CIFAR-10 under Krum and Trim attacks, it delivered over 50% improvement over FLTrust. The method was also validated on a real-world safari dataset for distributed wildlife monitoring, demonstrating its potential in socially impactful, safety-critical applications. The paper presents an 8-page study with four figures, contributing to the growing field of robust, privacy-preserving machine learning without the need for direct client data access.
- FoggyTrust localizes trust computation to fog nodes for better handling of heterogeneous data across clients.
- Achieves over 50% improvement over FLTrust on CIFAR-10 under Krum and Trim attacks.
- Combines local trust-based aggregation with optimizers like FedAdam and SCAFFOLD to reduce client drift.
Why It Matters
Enables secure collaborative ML in sensitive domains like wildlife monitoring without compromising data privacy or model accuracy.