With the rapidly increasing number of Android applications(apps), there are huge risks lurking in the app marketplaces as malicious software attempt to collect user privacy information without users' awareness. Once the user has installed malware, they are subject to privacy breach and information leakage. To tackle this problem, Android platform presents a list of permissions declared by apps to users and request users' authorization before installation. However, users with poor professional knowledge are very often not able to assess the risk of an app based on the only permission list. To allow better user assessment, an app description should reflect the application permissions involved. In our paper, we proposed a framework called AC-net that combines NLP and deep neural network to assess the consistency between mobile app descriptions and permissions. We paid more attention to the degree of description-to-permission fidelity instead of outputting only a simple answer. This paper is in the course of contribution, so the code will be posted later. The manually annotated data is in the folder of Dataset.