Privacy-preserving deep learning algorithm for big personal data analysis

For privacy-preserving analysing of big data, a deep learning method is proposed. The method transforms the sensitive part of the personal information into non-sensitive data. To implement this process, two-stage architecture is proposed. The modified sparse denoising autoencoder and CNN models have been used in the architecture. Modified sparse denoising autoencoder performs transformation of data and CNN classifies the transformed data. In order to achieve low loss in data transformation, the sparsification parameter is added to the objective function of the autoencoder by the Kullback-Leibler divergence function. Here, the efficiency evaluation of the model is conducted by the MSE (mean squared error) loss function. In order to evaluate the accuracy of the transformation process, the features derived from the sparse denoising autoencoder algorithm fed to the input of the deep CNN algorithm and the classification of the reconstructed data is classified to the Black (0), White (1) and Gray (2) classes. Since here conducted the transformation of the Black class data to the Gray class data, in the classification stage, the CNN algorithm is classified the Black class data as the Gray class with 0.99 accuracy. The comparison of the proposed method with simple autoencoder is provided and experiments conducted on Cleveland medical dataset extracted from the Heart Disease dataset, Arrhythmia and Skoda datasets showed that the proposed method outperforms other conventional methods.