Researchers propose sample selection methods using multitask autoencoders to improve federated learning with noisy data

A study presents novel sample selection methods for federated learning (a machine learning approach where multiple devices collaboratively train a model while preserving data privacy) that employ a multitask autoencoder to estimate sample contributions and filter noisy samples using unsupervised outlier detection techniques including one-class support vector machine (OCSVM), isolation forest (IF), and adaptive loss threshold (AT).

Validation on CIFAR10 and MNIST datasets across varying numbers of clients and non-IID distributions (non-independently distributed data) with noise levels up to 40% showed accuracy improvements of up to 7.02% on CIFAR10 with OCSVM and 1.83% on MNIST with AT using loss-based sample selection, with feature-based sample selection yielding additional gains of up to 0.99% on CIFAR10 with OCSVM.

The methods address a key challenge in federated learning where redundant, malicious, or abnormal samples lead to model degradation and inefficiency by using a central server to manage filtering and enhance feature-based sample selection through multi-class deep support vector data description (SVDD) loss.