Data Shift Under Delayed Labeling in Multi-Model Federated Learning
Cláudio G. S. Capanema, Fabrício Aguiar Silva, Leandro Aparecido Villas, Antônio A. F. Loureiro · 2025
The widespread use of various sensors (e.g., Global Navigation Satellite System (GNSS), cameras, accelerometers) has enabled devices, such as smartphones and vehicles, learn and automate numerous tasks (e.g., Human Activity Recognition, object recognition, driver monitoring, etc.). To support the learning of these tasks under privacy-preserving, scalable, and low communication cost constraints, Multi-model Federated Learning (MEFL) was recently introduced. This paradigm allows for the training of multiple independent models across federated clients, coordinated by a central server. A critical challenge in this context is data shift, where changes in data distribution can negatively affect model performance. Additionally, detecting data shift in a supervised manner presents challenges, as acquiring labeled data may be costly in certain contexts. To address this, we propose MultiFedPredict, a plugin to support a variety of non-IID data distributions and tackle the issue of late labeling, where clients delay detecting data drifts due to the cost associated with frequently acquiring labeled samples. Results show that our method can better deal with different data distributions and reduce the impact of data shift under delayed labeling specially on less heterogeneous data.