Software vulnerabilities in TensorFlow-based deep learning applications

Abstract

Usage of Deep Learning (DL) methods is ubiquitous. It is common in the DL/Artificial Intelligence domain to use 3rd party software. TensorFlow is one of the most popular Machine Learning (ML) platforms. Every software product is a subject to security failures which often result from software vulnerabilities. In this paper, we focus on threats related to 6 common types of threats in TensorFlow implementation. We identify them using Common Weakness Enumeration. We analyze more than 100 vulnerability instances. We focus on vulnerabilities’ severity, impact on confidentiality, integrity and availability, as well as possible results of exploitation. We also use Orthogonal Defect Classification (ODC). The results show that a majority of vulnerabilities are caused by missing/incorrect checking statements, however some fixes require more advanced algorithmic changes. Static Analysis Tools tested in our study show low effectiveness in detecting known vulnerabilities in TensorFlow, but we provide some recommendations based on the obtained alerts to improve overall code quality. Further analysis of vulnerabilities helped us to understand and characterize different vulnerability types and provide a set of observations. We believe that these observations can be useful for the creators of new static analysis tools as a source of inspiration and to build the test cases. We also aim to draw the programmers’ attention to the prevalence of vulnerabilities in deep learning applications and a low effectiveness of automatic tools to find software vulnerabilities in such products.