EASEAndroid: Automatic Policy Analysis and Refinement for Security Enhanced Android via Large-Scale Semi-Supervised Learning

Mandatory protection systems such as SELinux and SEAndroid harden operating system integrity. Unfortunately, policy development is error prone and requires lengthy refinement using audit logs from deployed systems. While prior work has studied SELinux policy in detail, SEAndroid is relatively new and has received little attention. SEAndroid policy engineering differs significantly from SELinux: Android fundamentally differs from traditional Linux; the same policy is used on millions of devices for which new audit logs are continually available; and audit logs contain a mix of benign and malicious accesses. In this paper, we propose EASEAndroid, the first SEAndroid analytic platform for automatic policy analysis and refinement. Our key insight is that the policy refinement process can be modeled and automated using semi-supervised learning. Given an existing policy and a small set of known access patterns, EASEAndroid continually expands the knowledge base as new audit logs become available, producing suggestions for policy refinement. We evaluate EASEAndroid on 1.3 million audit logs from real-world devices. EASEAndroid successfully learns 2,518 new access patterns and generates 331 new policy rules. During this process, EASEAndroid discovers eight categories of attack access patterns in real devices, two of which are new attacks directly against the SEAndroid MAC mechanism.