PAC: Privacy-preserving arrhythmia classification with neural networks

Recent advances in information technology such as the Internet of Things enable businesses and organisations to collect large amounts of data and apply advanced machine learning techniques in order to infer valuable insights and improve predictions. Unfortunately, such benets come with a high cost in terms of privacy exposures given

the high sensitivity of the data that are usually analysed/processed at third party servers. In this study, we aim at protecting health data, more precisely, Electro-Cardiogram (ECG) data while enabling an efficient prediction of arrhythmia using neural networks. We propose a solution named PAC that combines the use of Neural Networks with

secure two-party computation. To achieve a good trade-off between privacy, accuracy, and efficiency, we rst build a dedicated, efficient neural network model for which the underlying operations can be further performed while data is privacy protected. The resulting model consists of two fully connected layers which perform small size matrix multiplication and one activation layer which executes a square function. The solution

is implemented using the ABY framework and makes use of Arithmetic and Boolean circuits. Several approximations are performed over the representation of the data, accordingly. PAC also supports classications in batches when needed. Experimental results run on the PhysioBank datasets show an accuracy of 96.34% which outperforms existing solutions.