Intel Knights Landing Processors have shared last level cache (LLC) across all the tiles using MESIF protocol and uses a mesh network of Caching and Homing Agents(CHA)s. Due to the structure of the network, the cache access is non uniform in nature having significant difference in cache hit times. In this paper, we try to exploit this idea to leak secret from a victim process. First, we show a naive implementation of the attack using a gem5 simulator that achieves 100\% accuracy of extracting the secret bits. Then we replicate the attack in a Intel Xeon Phi 7290@ 1.50 GHz Knight's Landing CPU to show the efficacy of the attack. In real machine we can leak the secret from a victim process at 85\% accuracy and ~350 kbps bandwidth. All the attacks were done on a machine without any root or sudo privileges, so this shows the strength of the attack. This can be further extended to leak secrets from different processes given the vulnerable patterns may exist in many libraries. Other processors with similar architecture (last level distributed cache in mesh networks) can also be vulnerable to similar attack strategy.
翻译:英特尔骑士着陆处理器在使用 MESIF 协议的所有瓷砖中共享最后一级缓存( LLC), 并使用 Caching 和 Homing Agents (CHA) 的网状网络网络网络。 由于网络的结构, 缓存存访问在性质上是非统一的, 在缓存点击时间上差异很大。 在本文中, 我们试图利用这个想法从受害者程序中泄露秘密。 首先, 我们展示了使用宝石5 模拟器进行攻击的天真性, 其精确度达到了100 ⁇ 提取秘密部分的精度。 然后, 我们复制了在 Intel Xeon Phi 7290@ 1. 150 GHMz Knight 登陆 CPU 中进行攻击, 以显示袭击的功效。 在真实机器中, 我们可以以 85 ⁇ 精确度 和 ~ 350 kbps 带宽度向受害者进程泄露秘密。 所有攻击都是在没有根或sudo 特权的情况下进行的, 这表明攻击的强度 。 这可以进一步扩展到由于许多图书馆的脆弱模式可能存在不同过程的机密。 其他处理器( 类似攻击网络中的最后一级分配脆弱缓冲隐藏) 也可以成为类似的战略 。