Federated personalized random forest for human activity recognition.

Area of Science:

• Machine Learning
• Data Privacy
• Human Activity Recognition

Background:

• User data is often siloed due to privacy regulations like GDPR, hindering centralized machine learning model training.
• Federated learning allows collaborative model training without exposing original data, with random forests being a popular choice due to speed and accuracy.
• Existing federated random forest models lack personalization for specific tasks like human activity recognition.

Purpose of the Study:

• To propose a privacy-protected federated personalized random forest framework for human activity recognition.
• To enable personalized machine learning services by training models only with similar users.
• To enhance user privacy during the federated training process.

Main Methods:

• Developed a federated personalized random forest framework incorporating locality-sensitive hashing for user similarity calculation.
• Implemented an incremental model selection strategy based on ensemble learning characteristics for personalized training.
• Integrated differential privacy mechanisms to protect user data during the training stage.

Main Results:

• Experimental validation on human activity recognition datasets demonstrates the framework's effectiveness.
• The proposed method achieves personalized model training while preserving user privacy.
• Federated learning with personalized random forests shows promise for activity recognition tasks.

Conclusions:

• The privacy-protected federated personalized random forest framework effectively addresses the need for personalized human activity recognition.
• Locality-sensitive hashing and differential privacy are key components for achieving privacy-preserving personalization in federated learning.
• This approach facilitates the development of tailored machine learning applications in scenarios with sensitive user data.