Low-Round Federated Learning via Feature Synthesis and Hard-Sample-Driven Enhancement
Abstract
This project uses three publicly available datasets, CIFAR-10, CIFAR-100, and SVHN, to experimentally validate the FedSynth2 federated learning algorithm. Among them, CIFAR-10 contains 10 categories and a total of 60000 32 × 32 color images. CIFAR-100 can be expanded to 100 categories under the same data scale, which can reflect more complex fine-grained classification scenarios; SVHN is derived from real street view digital images and has a strong practical application background. The three types of datasets cover basic image classification, fine-grained visual recognition, and real-world digital recognition tasks, and can evaluate the algorithm's generalization ability and applicability from different perspectives. To simulate the heterogeneous features of data in federated learning, the experiment used Dirichlet distribution to partition the training data into client and controlled the Non IID degree by adjusting the parameter α. The focus was on the strong heterogeneous scenario under the condition of α=0.05 to verify the convergence performance and stability of FedSynth2 under limited communication epochs and complex data distribution conditions. All experimental data are sourced from publicly available datasets and do not involve personal privacy information. They can be used for academic research and algorithm evaluation.
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.