Federated transfer learning for auxiliary classifier generative adversarial networks: framework and industrial...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Data privacy concerns hinder collaborative model training in industrial settings.
• Decentralized data islands prevent the creation of effective personalized models.
• Existing methods struggle to balance personalization with strict data privacy requirements.

Purpose of the Study:

• To propose a novel Federated Transfer Learning (FL) framework for privacy-preserving personalized machine learning in manufacturing.
• To address the challenge of training personalized models on isolated industrial datasets.
• To enhance the performance of machine learning models in real-world industrial scenarios.

Main Methods:

• Developed a Federated Transfer Learning (FL) framework integrating Auxiliary Classifier Generative Adversarial Networks (ACGAN).
• Utilized Federated Learning (FL) for privacy-preserving global model training on decentralized data.
• Employed Transfer Learning (TL) to adapt the global model for personalized applications.
• Leveraged ACGAN to generate synthetic data, bridging FL and TL while maintaining privacy.

Main Results:

• The ACGAN-FTL framework achieved 0.81 accuracy, 0.86 precision, 0.74 recall, and 0.79 F1 score in a carbon anode quality prediction task.
• Demonstrated significant performance gains compared to a baseline method without FL and TL, with increases of 13% (accuracy), 11% (precision), 16% (recall), and 15% (F1 score).
• Ensured data privacy throughout the entire machine learning process.

Conclusions:

• The proposed ACGAN-FTL framework effectively enables privacy-preserving personalized machine learning in industrial settings.
• The framework successfully addresses the limitations of isolated data islands and privacy concerns.
• ACGAN-FTL meets the performance requirements for real-world industrial applications, offering substantial improvements over traditional methods.