Secure Image Reconstruction using Deep Learning-based Autoencoder with Integrated Encryption Layers

Abstract

This study presents an autoencoder model designed for secure image reconstruction through the integration of encryption and decryption layers within its framework. The major goal is to achieve more effective image reconstruction while safeguarding data integrity. A convolutional neural network (CNN) is first utilized as the primary architecture, attaining a reconstruction accuracy of 90.63% with 2.3737x losses. This brought an opportunity for further improvement, and thus we propose the improved model with the integration of CNN and bidirectional gated recurrent unit (BiGRU) as hybrid model. The integration of CNN-BiGRU leverages the feature extraction advantage of CNN and the temporal processing ability of BiGRU to a great improvement of reconstruction accuracy, reaching 95.57% and validation accuracy stabilizing around 0.85 at the end of training. The model exhibits great accuracy without significant overfitting, thus acquiring robust characteristics crucial for precise image reconstruction. In this work, the hybrid model outperforms the conventional CNN-only architectures for secure image reconstruction and can thus be considered a potential approach when high fidelity with security is required in processing image data.

Keywords

• Autoencoder (AE)
• Convolutional neural network (CNN)
• Bidirectional gated recurrent neural network (BiGRU)
• Deep learning (DL)