Learning Security Policies

Óbudai Egyetem

Alkalmazott Informatikai és Alkalmazott Matematikai Doktori Iskola

Csiszár Orsolya

informatikai tudományok

Access control systems, such as an in sandboxing software or mandatory access control systems (e.g. SELinux) require a policy. The policy dictates which operations are to be allowed and which are to be forbidden. Designing a policy, especially a system-wide one, is a complex and error-prone process. Auto-generation tools exist (such as audit2allow for SELinux). Often these tools take the system logs as a source of information about the permissions that the systems need to operate normally. However, these logs only reflect the permissions that have so far been requested. As such, their completeness depends on the degree to which the previous operation is an indicator of what permissions will be needed in the future – that is, their quality depends on the quality of the test cases. Moreover, tools that auto-generate policy often over approximate the required permissions, thereby weakening the security of the system. For instance, the aa-genprof tool that is part of the AppArmor toolchain detects if file access is below a user’s home directory and will propose to enable access to the corresponding location below any user’s home directory. Whether this is the right decision depends on the specific software and file access that is requested. A more intelligent solution that uses context information to predict future access patterns is desirable.
The purpose of this research topic is to determine how such policies can be learned and how existing/auto-generated policies can be cross-checked against a declarative statement of the administrator’s intent – a “red line” that must not be crossed. Since data samples are small and traditional machine learning approaches, such as neural nets, are therefore difficult to deploy, we propose to use inductive logic programming (ILP) to learn policies in this context.

1 fő

Óbudai Egyetem

2022-07-01

angol