Enhancing privacy-preserving brain tumor classification with adaptive reputation-aware federated learning and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Privacy

Background:

• Automated brain tumor diagnosis via MRI is crucial but hindered by data privacy and scarcity.
• Federated learning (FL) offers a solution by enabling collaborative training without raw data sharing, yet faces its own challenges.
• Existing FL methods struggle with data heterogeneity and privacy vulnerabilities like model inversion attacks.

Purpose of the Study:

• To introduce Federated Adaptive Reputation-aware aggregation with CKKS Homomorphic encryption (FedARCH), a novel FL framework for cross-silo medical image analysis.
• To enhance global model accuracy and robustness against noisy data and adversarial attacks.
• To ensure data privacy through efficient homomorphic encryption during federated model aggregation.

Main Methods:

• Developed FedARCH, a federated learning framework employing reputation scores for weighted client aggregation.
• Integrated CKKS homomorphic encryption for secure, privacy-preserving operations on encrypted model weights.
• Implemented dynamic performance management using smoothing and decay factors for adaptive aggregation.

Main Results:

• FedARCH achieved a high accuracy of 99.39% in distinguishing brain tumor classes.
• The framework maintained 94% accuracy with 50% noisy clients, significantly outperforming standard FL (33% accuracy).
• Security analysis confirmed FedARCH's effectiveness in mitigating privacy risks and computational overhead.

Conclusions:

• FedARCH offers a robust and privacy-preserving solution for federated medical image analysis.
• The proposed reputation-aware aggregation and homomorphic encryption effectively address key challenges in FL for healthcare.
• FedARCH demonstrates significant potential for improving brain tumor diagnosis accuracy and reliability in real-world federated settings.