A Light-Weight Crowdsourcing Aggregation in Privacy-Preserving Federated Learning System
Ke Zhang, Siu Ming Yiu, Lucas C. K. Hui · 2020
Federated Machine Learning (FML) sheds light on secure distributed machine learning. However, generic FML methods may lead to privacy-leakage through the sharing of training information of individual models and have relatively poor performance when the training datasets for individual models are biased and diversified. This is a problem in combining models trained in different scenarios of IoT devices since the available training datasets are usually limited and biased. To tackle this problem, we propose a novel approach to precisely ensemble results from different models in distributed edge devices. Instead of passing the training information of individual models around that requires a relatively large amount of bandwidth and compromises data privacy, we suggest employing a trusted central agent that only collects different inference results from edge devices. Then based on a limited amount of labeled data, the agent runs a designed statistical iterative crowdsourcing algorithm to combine results for a more accurate aggregated prediction towards a user query. Our proposed system model, "Privacy-Preserving Federated Learning System", together with our light-weight Secure Crowdsourcing Aggregation (SC-Agg) algorithm, provide a more accurate prediction for outside queries at little cost without any prior knowledge of what query will be submitted. We experimentally verify that in our system, SC-Agg consistently outperforms the majority voting method and the best performing model of the ensemble in all testing scenarios. We believe that SC-Agg fits the real-world IoT applications better than other methods, such as the vanilla majority voting, for its robustness and better performance.