Safety or Not? A Comparative Study for Deep Learning Apps on Smartphones

Recent years have witnessed an astonishing explosion in the evolution of mobile applications powered by deep learning (DL) technologies. Considering that inference of DL models in the cloud requires transferring user data to server, which is prone to the risk of user privacy leakage, many developers choose to deploy the models on local devices for executing the inference process. However, this also raises a number of other security issues, such as adversarial attacks, model stealing attacks, etc. To explore the security issues that exist in deep learning applications (DL apps), we conducted the first comprehensive comparative study of the top 200 apps in each category on Google Play. We built DLApplnspector, a vulnerability detection tool that combines dynamic and static analysis methods for dissecting apps, which helped us automate our empirical study on real-world mobile DL apps. First, we identify DL apps and extract their models by using DL Checker. Subsequently, Static Scoper is provided to detect encryption and reusability of DL models. Finally, within Dynamic Scoper, we use reverse engineering techniques on the network traffic to parse out the packets and collect side-channel information during application runtime. Our research shows that the majority of developers prefer to use open-source models, with almost 92% of models successfully parsed. This suggests that most models are unprotected. DL apps are more likely to upload user behaviour and collect private data than Non-DL apps. We provide security recommendations for developers and users to address the issues discovered.