Federated Learning With Differential Privacy: A Production Architecture
Meta has been evaluating federated learning with differential privacy (FL-DP) as a privacy-enhancing technology for its products. The approach offers two distinct privacy benefits: it keeps user data on mobile devices by training models in a distributed fashion, and it introduces noise that reduces the risk of a model memorizing individual data points.
While the concept is straightforward, building production systems around FL-DP presents problems that conventional machine learning tools and workflows don't address. Meta's engineering team has developed a system architecture to tackle these issues, aiming to maintain user privacy without sacrificing the quality of intelligent, interactive features.
The Core Challenge
In an FL-DP setup, mobile devices handle training locally, and only aggregated updates—after noise has been applied via differential privacy—reach the global model. DP is considered the strongest available defense against model memorization, even under extreme threat models like reconstruction attacks.
But moving to this decentralized paradigm introduces several practical obstacles:
- Label balancing, feature normalization, and metrics calculation become difficult when data can't be directly inspected.
- Mobile release cycles are slower than back-end server cycles, slowing iteration.
- Training speed suffers because computation is distributed across mobile devices.
- System logging must be anonymized to preserve privacy.
Design for Hybrid Data and Fast Iteration
Meta's proposed architecture combines training on server-side user data with inferences that leverage device-side-only data—information that stays on the user's phone. The infrastructure spans three tiers: mobile devices, trusted execution environments, and conventional back-end servers.
Prior research has focused on the algorithmic side of efficient federated training, but less attention has been paid to the surrounding infrastructure needed for fast developer iteration and scalable deployment at inference time. A key goal of this design was closing that gap.
The system was validated using an in-house FL library that works across Meta's apps, including Facebook and Instagram. The library is designed to scale training to millions of devices and inference to billions. When benchmarked against conventional server-trained models, the federated approach showed minimal performance degradation while respecting the constraints of on-device compute, storage, and power.
Remaining Hurdles
Despite the architecture's success in training and deploying production FL models, developer speed remains the most significant barrier to scaling federated ML. Model iteration cycles in the federated environment are still orders of magnitude slower than equivalent work in a centralized system. The team plans to continue refining the architecture to address this bottleneck.
