EAGER: Side Channels Go Deep - Leveraging Deep Learning for Side-channel Analysis and Protection

点击翻译按钮获取中文摘要

Side-channel attacks (SCAs) have presented serious threats to confidentiality and privacy in various areas including finance, transportation, mobile communications, and clouds. The new exploits, Meltdown and Spectre, have revealed that indispensable performance-related optimizations of modern computer architecture have turned into fundamental vulnerabilities for information leakage. However, finding SCA leakage thoroughly on real systems can be challenging, and inferior leakage evaluation methods used by the system developer would result in devices or software without appropriate protection entering field operations, vulnerable to dedicated adversaries possessing more sophisticated attacks. The recent advancement of machine learning techniques, particularly deep neural networks (DNNs), has facilitated SCAs to learn and utilize side-channel power leakage of complex forms, resulting in outperforming the strongest classic template attacks and even breaking certain common SCA countermeasures. Power leakage and security evaluation has shifted to DL-based methods. However, there is little application of DNNs in microarchitectural attacks, despite the surging discovery and exploitation of vulnerable microarchitectures. This project aims to leverage the rapidly evolving advances of deep learning in both microarchitectural SCAs and countermeasures. The novelties of the project lie in both a new microarchitecture monitor and the follow-on data analytic and system obfuscation methods. The project's broad significance and importance are it will advance the state-of-the-art on microarchitectural attacks, side-channel security evaluation, and protection against confidentiality and privacy breach.This project investigates foundational issues of applying deep learning techniques for both microarchitectural side-channel analysis and protection. The technical approach includes a persistent cache monitoring mechanism, which significantly improves the observability of the victim execution by the spy and captures detailed information leakage in timing traces. Appropriate DNN models are being built to exploit the timing traces for secret retrieval. The entire framework of microarchitectural monitoring and deep learning-based attacks is applicable to diverse platforms, enabled by cross-device transfer learning and generative adversarial networks (GANs). The concept of adversarial examples is being leveraged to direct novel effective countermeasures against DL-based side-channel attacks. The outcome of this project, thorough DL-based side-channel attacks, sound security evaluation, and efficient protections, will have profound impact in securing the clouds and critical systems and infrastructures.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.