Towards Automated Analysis of Microarchitectural Attacks using Machine Learning

Gulmezoglu, Berk

Etd

Towards Automated Analysis of Microarchitectural Attacks using Machine Learning

Public

Cloud computing has gained tremendous popularity among small and mid-size businesses, providing many companies to access cloud services without the need for investing in computer software and hardware. In order to o↵er the highest performance to clients, multi-core servers are mostly preferred by the cloud providers, which yields to sharing the same hardware resources among clients. Although it is expected that hardware virtualization is a sufficient protection against potential attackers from co-located users, microarchitectural attacks still pose a significant threat in the cloud due to the shared hardware resources. Moreover, similar attack techniques are applicable in both personal computers and mobile phones when a benign-looking malicious application is installed in the system. Today, microarchitectural attacks are more important threat than ever, since the capabilities of the attacks have been extended tremendously. In order to recover confidential information from a third-party by implementing a microarchitectural attack, thousands or millions of side-channel measurements are collected. Since the side-channel analysis requires engineering expertise to extract the information such as handling mis-alignment and extracting the secret bits one by one, it takes huge amount of time to process millions of side-channel data samples. Thanks to recent advances in Machine Learning, the complex tasks can be handled efficiently with the high computation power of GPUs. The linear and non-linear problem solving capabilities of Machine Learning algorithms are integrated with giant matrix multiplications, the success rate improves drastically. However, the appropriate application of Machine Learning techniques on side-channel measurements is still an ongoing research area, which can provide huge time and performance gain in terms of leakage extraction. The goal of this dissertation is to propose Machine Learning based approaches to processing of side-channel measurements in different platforms. First, we introduce a targeted co-location technique by using cryptographic libraries on public clouds. Then, a full RSA key recovery attack based on last-level cache is demonstrated on Amazon EC2 public cloud. Next, we implement a Machine Learning assisted cache attack on a public cloud as well as showing that ping requests can be used to identify the co-located VMs. Furthermore, we show that privacy of personal computer and mobile phone users can be violated by third-party applications through microarchitectural attacks. For this purpose, we detect the visited websites, launched applications and watched trailers, as well as, comparing the performance of several Machine Learning techniques. Finally, we propose a Recurrent Neural Network based unsupervised detection mechanism for microarchitectural attacks. We achieve a low false alarm rate and performance overhead by combining the sequence learning capabilities of RNNs and computation power of GPUs.

Creator