Analysis of Privacy-Enhancing Technologies in Open-Source Federated Learning Frameworks for Driver Activity...

Area of Science:

• Computer Science
• Cybersecurity
• Human-Computer Interaction

Background:

• Wearable devices and smartphones collect sensitive driver data (audio, video, location, health).
• Processing this data requires strict adherence to personal data security and privacy regulations.
• Federated learning (FL) is a privacy-preserving paradigm, but lacks formal privacy guarantees and is vulnerable to attacks.

Purpose of the Study:

• Analyze privacy-preserving techniques for FL in driver monitoring.
• Compare implementations in open-source FL frameworks.
• Evaluate the impact of privacy techniques on FL training efficiency and accuracy for driver behavior analysis.

Main Methods:

• Comparative review and analysis of privacy-preserving techniques in open-source FL frameworks.
• Evaluation of technique impact on global model accuracy, training time, and network traffic.
• Experimental setup focused on driver activity monitoring using smartphone sensor data.

Main Results:

• Current privacy-preserving techniques in open-source FL frameworks have limitations.
• These limitations significantly impact the practical application of FL for driver monitoring.
• The overhead of privacy techniques affects training efficiency and resource utilization.

Conclusions:

• Existing privacy techniques in FL frameworks are not yet fully suitable for real-time driver monitoring applications.
• Further research is needed to enhance the efficiency and security of FL for sensitive data processing.
• The practical application of FL is currently restricted to cross-silo settings due to these limitations.