Private Federated Learning In Real World Application -- A Case Study

TL;DR

This work tackles privacy-preserving training of an app-selection predictor via Private Federated Learning on edge devices, addressing user data privacy while enabling on-device adaptation. It introduces a DNN architecture with cross-entity attention and explicit uncertainty handling, and integrates differential privacy with federated aggregation, evaluated through offline simulations and on-device experiments. The results demonstrate that PFL can adapt to evolving user behavior with only modest accuracy degradation due to privacy constraints, and that fine-tuning from existing checkpoints often outperforms training from scratch; on-device plugins enable secure data processing and server-side aggregation. The study provides practical guidance for deploying PFL in industry, including data-generation standardization, plugin design, and hyper-parameter considerations, and outlines future work on privacy–utility optimization and extending PFL to broader predictive tasks.

Abstract

This paper presents an implementation of machine learning model training using private federated learning (PFL) on edge devices. We introduce a novel framework that uses PFL to address the challenge of training a model using users' private data. The framework ensures that user data remain on individual devices, with only essential model updates transmitted to a central server for aggregation with privacy guarantees. We detail the architecture of our app selection model, which incorporates a neural network with attention mechanisms and ambiguity handling through uncertainty management. Experiments conducted through off-line simulations and on device training demonstrate the feasibility of our approach in real-world scenarios. Our results show the potential of PFL to improve the accuracy of an app selection model by adapting to changes in user behavior over time, while adhering to privacy standards. The insights gained from this study are important for industries looking to implement PFL, offering a robust strategy for training a predictive model directly on edge devices while ensuring user data privacy.

Figures (4)

• Figure 1: Model architecture
• Figure 2: Fine-tuning Cymba from an existing checkpoint
• Figure 3: Using simulations to predict the performance of the fine-tuning of Top Layers of the Cymba model as a function of Number of devices (cohort size) and corresponding total number of data points.
• Figure 4: PFL on device training workflow