Multi-Source Selection Transfer Learning with Privacy-Preserving | JoVE Visualize

Area of Science:

Machine Learning
Artificial Intelligence
Data Privacy

Background:

Transfer learning leverages source domain knowledge for target tasks, but privacy concerns and domain distribution differences hinder its application.
Existing privacy-preserving transfer learning methods often neglect data distribution variations, leading to negative transfer and reduced performance.
Research on privacy protection within transfer learning is limited, especially concerning multi-source domain scenarios.

Purpose of the Study:

To propose a privacy-preserving multi-source transfer learning algorithm (MultiSTLP) for scenarios with unlabeled target data and multiple labeled source domains.
To address the challenges of marginal and conditional probability distribution differences between domains in a privacy-conscious manner.
To enhance classification accuracy and efficiency in transfer learning by integrating multi-source learning and group probability concepts.

Main Methods:

Developed a multi-source selection transfer learning algorithm (MultiSTLP) incorporating privacy-preserving techniques.
Fused multi-source transfer learning and group probability concepts into a support vector machine (SVM) framework.
Implemented a source domain selection mechanism to improve computational efficiency by identifying representative datasets.

Main Results:

MultiSTLP effectively adapts to marginal and conditional probability distribution differences between source and target domains.
The algorithm successfully protects the privacy of target data while enhancing classification accuracy.
Experimental results on real datasets demonstrate the effectiveness and advantages of MultiSTLP over state-of-the-art transfer learning algorithms.

Conclusions:

MultiSTLP offers a robust solution for privacy-preserving transfer learning in complex multi-source scenarios.
The proposed method enhances classification performance by effectively managing domain distribution discrepancies and leveraging group probability information.
This work advances the field by providing an efficient and privacy-conscious approach to transfer learning with unlabeled target data.