Approximation-based energy-efficient cyber-secured image classification framework

Abstract

In this work, an energy-efficient cyber-secured framework for deep learning-based image classification is proposed. This simultaneously addresses two major concerns in relevant applications, which are typically handled separately in the existing works. An image approximation-based data storage scheme to improve the efficiency of memory usage while reducing energy consumption at both the source and user ends is discussed. Also, the proposed framework mitigates the impacts of two different adversarial attacks, notably retaining performance. The experimental analysis signifies the academic and industrial importance of this work as it demonstrates reductions of 62.5% in energy consumption for image classification when accessing memory and in the effective memory sizes of both ends by the same amount. During the improvement of memory efficiency, the multi-scale structural similarity index measure (MS-SSIM) is found to be the optimum image quality assessment method among different similarity-based metrics for the image classification task with approximated images and an average image quality of 0.9449 in terms of MS-SSIM is maintained. Also, a comparative analysis of three different classifiers with different depths indicates that the proposed scheme maintains up to 90.17% of original classification accuracy under normal and cyber-attack scenarios, effectively defending against untargeted and targeted white-box adversarial attacks with varying parameters.