PT-Rand: Practical Mitigation of Data-only Attacks against Page Tables
Download: Paper (PDF)
Date: 27 Feb 2017
Document Type: Reports
Associated Event: NDSS Symposium 2017
Kernel exploits constitute a powerful attack class allowing attackers to gain full control over a system. Various kernel hardening solutions have been proposed or deployed in practice to protect the kernel against code injection (e.g., DEP) or code-reuse exploits (e.g., CFI). However, the security of all these hardening techniques relies heavily on the assumption that kernel page tables cannot be manipulated, e.g., by means of dataonly attacks. Ensuring kernel page tables integrity is not only essential for kernel security but also a challenging task in practice since existing solutions require hardware trust anchors, costly hypervisors, or inefficient integrity checks.
In this paper, we first motivate the importance of protecting kernel page tables by presenting a data-only attack against page tables to bypass the recently released CFI-based (Linux) kernel hardening technique RAP. Thereafter, we present the design and implementation of PT-Rand, the first practical solution to protect kernel page tables that does not suffer from the mentioned deficiencies of previous proposals. PT-Rand randomizes the location of page tables and tackles a number of challenges to ensure that the location of page tables is not leaked. This effectively prevents the attacker from manipulating access permissions of code pages, thereby enabling secure enforcement of kernel exploit mitigation technologies such as CFI. We extensively evaluate our prototype implementation of PT-Rand for the current Linux kernel on the popular Linux distribution Debian and report a low overhead of 0.22% for common benchmarks. Moreover, we combine RAP with PT-Rand to protect RAP against data-only attacks on kernel page tables.