When Mobile Apps Going Deep: An Empirical Study of Mobile Deep Learning

Deep learning (DL) is a game-changing technique in mobile scenarios, as already proven by the academic community. However, no prior literature has studied the adoption of DL in the mobile wild. To fill such gap, in this work, we carry out the first empirical study to demystify how DL is utilized in mobile apps. Based on static analysis technique, we first build a framework that can help accurately identify the apps with DL embedded and extract the DL models from those apps. We then perform comprehensive and in-depth analysis into those apps and models, and make interesting and valuable findings out of the analysis results. As one of the key observations, we find that DL is becoming increasingly popular on mobile apps, and the roles played by DL are mostly critical rather than dispensable. On the other side, however, the potential of DL is far from being fully utilized, as we observe that most in-the-wild DL models are quite lightweight and not well optimized. Our work also provides useful implications for researchers and developers on the related fields.