Encrypted federated learning for secure decentralized collaboration in cancer image analysis

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Imaging

Background:

• AI in cancer research is limited by data access due to privacy regulations.
• Federated and swarm learning offer decentralized training but risk data breaches via weight updates.
• Untrusted servers can potentially compromise privacy through model inversion or membership inference attacks.

Purpose of the Study:

• To demonstrate the first successful implementation of somewhat-homomorphically-encrypted federated learning (SHEFL) for cancer image analysis.
• To evaluate the performance of SHEFL against local and central training methods.
• To showcase SHEFL's potential for secure, collaborative AI model development across institutions.

Main Methods:

• Implementation of SHEFL for training AI models on multicentric datasets.
• Application of SHEFL to clinically relevant tasks in radiology and histopathology cancer image analysis.
• Encrypted weight transfer and in-space model updates to preserve data privacy.

Main Results:

• SHEFL successfully enabled AI model training on multicentric cancer image datasets.
• AI models trained with SHEFL outperformed models trained locally.
• SHEFL-trained models performed comparably to centrally trained models.

Conclusions:

• SHEFL provides a viable solution for privacy-preserving collaborative AI training in oncology.
• This approach allows multiple institutions to co-train AI models without compromising data governance.
• SHEFL eliminates the transmission of decryptable data to untrusted servers, enhancing security.