PPM: A Privacy-Preserving Framework for Intelligent Medical Diagnosis Systems

The extensive utilization of deep learning techniques has significantly boosted the efficacy of intelligent Computer-Aided Diagnosis (CAD) using medical data from multiple modalities. Deep learning techniques generally suffer from high computational complexity and cloud computing can effectively empower deep learning techniques by optimized software (e.g. distributed training/inference) and specific hardware (e.g. GPUs, FPGAs and TPUs). However, cloud computing also increases the risk of medical data leakage, which greatly hurts patients’ privacy. There has been several techniques including homomorphic encryption approach proposed to preserve patients’ privacy. However, none of them can be directly applied to deep neural networks to improve medical CAD applications via multi-modal data fusion. In this paper, we propose PPM, a framework integrating semantic level multi-modal data fusion neural network with homomorphic encryption mechanism, in order to preserve patients’ privacy while maintaining the diagnostic performance. We conduct experiments on multi-source and multi-modal medical data for Upper Gastrointestinal (UGI) cancer screening.