Inspecting the Running Process of Horizontal Federated Learning via Visual Analytics

Area of Science:

• Machine Learning
• Data Visualization
• Privacy-Preserving Technologies

Background:

• Horizontal Federated Learning (HFL) enables collaborative model training while preserving data privacy.
• Current HFL inspection methods offer limited insights into client behavior and model contributions.
• Existing visualization tools provide basic performance dashboards but lack in-depth analytical capabilities.

Purpose of the Study:

• To design an exploratory visual analytics system, HFLens, for inspecting the Horizontal Federated Learning process.
• To enable detailed analysis of client behavior, model performance, and contributions within HFL.
• To facilitate anomaly detection and intervention strategies in privacy-preserving machine learning.

Main Methods:

• Development of HFLens, a visual analytics system adhering to HFL privacy protocols.
• Implementation of comparative visual interpretation at overview, communication round, and client instance levels.
• Facilitation of correlation analysis, anomaly identification, and client contribution assessment.

Main Results:

• HFLens supports intensive exploration of the HFL process, going beyond shallow-level analysis.
• The system enables comparative interpretation across different levels of granularity (overview, round, client).
• Case studies and expert feedback confirm HFLens' efficacy in understanding and diagnosing HFL processes.

Conclusions:

• HFLens enhances the interpretability and diagnosability of Horizontal Federated Learning.
• The system empowers users to investigate client behaviors, assess contributions, and identify anomalies.
• HFLens represents a significant advancement in visualizing and analyzing privacy-preserving federated learning.