Micro-architectural Threats to Modern Computing Systems

With the abundance of cheap computing power and high-speed internet, cloud and mobile computing replaced traditional computers. As computing models evolved, newer CPUs were fitted with additional cores and larger caches to accommodate run multiple processes concurrently. In direct relation to these changes, shared hardware resources emerged and became a source of side-channel leakage. Although side-channel attacks have been known for a long time, these changes made them practical on shared hardware systems. In addition to side-channels, concurrent execution also opened the door to practical quality of service attacks (QoS). The goal of this dissertation is to identify side-channel leakages and architectural bottlenecks on modern computing systems and introduce exploits. To that end, we introduce side-channel attacks on cloud systems to recover sensitive information such as code execution, software identity as well as cryptographic secrets. Moreover, we introduce a hard to detect QoS attack that can cause over 90+\% slowdown. We demonstrate our attack by designing an Android app that causes degradation via memory bus locking. While practical and quite powerful, mounting side-channel attacks is akin to listening on a private conversation in a crowded train station. Significant manual labor is required to de-noise and synchronizes the leakage trace and extract features. With this motivation, we apply machine learning (ML) to automate and scale the data analysis. We show that classical machine learning methods, as well as more complicated convolutional neural networks (CNN), can be trained to extract useful information from side-channel leakage trace. Finally, we propose the DeepCloak framework as a countermeasure against side-channel attacks. We argue that by exploiting adversarial learning (AL), an inherent weakness of ML, as a defensive tool against side-channel attacks, we can cloak side-channel trace of a process. With DeepCloak, we show that it is possible to trick highly accurate (99+\% accuracy) CNN classifiers. Moreover, we investigate defenses against AL to determine if an attacker can protect itself from DeepCloak by applying adversarial re-training and defensive distillation. We show that even in the presence of an intelligent adversary that employs such techniques, DeepCloak still succeeds