An Edge-oriented Deep Learning Model Security Assessment Framework

As computer performance advances and deep learning models proliferate, their applications expand, but concerns arise due to their opaque nature and security issues. Edge intelligence requires more from these models but faces limitations like insufficient computing power and memory. Existing research on model compression mainly focuses on the trade-off between compression and accuracy, neglecting deployment at the edge and its impact on factors beyond accuracy. Furthermore, evaluation methods for model performance and robustness in compression techniques are underdeveloped. To address these issues, this paper introduces a comprehensive framework for evaluating edge intelligence deep learning models. It conducts a thorough assessment of these models on common image classification datasets, comparing their performance with three compression algorithms and analyzing their security implications. The proposed framework aims to provide a holistic assessment of deep learning models, promoting trustworthy edge intelligence, enabling secure deployments, and enhancing understanding of model decision-making processes.