Training Medical-Diagnosis Neural Networks on the Cloud with Privacy-Sensitive Patient Data from Multiple Clients

Area of Science:

• Artificial intelligence in medicine
• Computational neuroscience
• Data privacy and security

Background:

• Artificial neural networks (ANNs) offer transformative potential for medical diagnosis.
• Protecting patient data privacy during cloud-based model training is a significant challenge.
• Existing privacy methods like homomorphic encryption, differential privacy, and federated learning have limitations for outsourced cloud training.

Purpose of the Study:

• To propose a novel privacy-preserving method for outsourcing artificial neural network (ANN) model training to the cloud.
• To enable cloud-based training without requiring clients to perform local or synchronized operations.
• To maintain high model accuracy while safeguarding distributed patient data.

Main Methods:

• Implementation of a matrix masking technique for data encryption before cloud outsourcing.
• Development of a cloud-based training framework utilizing the masked data.
• Experimental validation using real-world datasets for Alzheimer's and Parkinson's disease diagnosis.

Main Results:

• Matrix masking allows complete outsourcing of model training operations to the cloud.
• The proposed method effectively protects patient data privacy during distributed training.
• ANN models trained using matrix-masked data achieve accuracy comparable to models trained on raw data.

Conclusions:

• Matrix masking presents a viable solution for privacy-preserving, outsourced cloud-based training of ANNs in medical diagnosis.
• This approach overcomes the limitations of traditional privacy-enhancing technologies for this specific application.
• The method demonstrates practical applicability in sensitive medical domains like neurodegenerative disease analysis.