[Paper review] SGX-Tor: A Secure and Practical Tor Anonymity Network With SGX Enclaves

S. Kim, J. Han, J. Ha, T. Kim, and D. Han, “SGX-TOR: A secure and practical TOR anonymity network with SGX enclaves”, IEEE/ACM Transactions on Networking, vol. 26,no. 5, pp. 2174–2187, 2018

Context

Intel SGX is a set of extensions to the Intel architecture that aims to provide integrity and confidentiality guarantees to security sensitive computation performed on a computer where all the privileged softwares (kernel, hypervisor, etc) are potentially malicious\(^1\). It provides isolated execution by putting and executing the code and data of an application inside a secure container called an enclave\(^2\). Indeed, by storing and computing information into enclaves, Intel opened opportunities for a wide set of usages. Using secure enclaves to protect sensitive information is of increasing interest to the community of researchers who see in this method a new approach to preserve anonymity of individuals, especially on cloud servers. Hence the interest of the authors of this paper in the implementation of this technology on the TOR network considered as being amongst the best way of preserving the anonymity of online users. Indeed, TOR is an anonymity network built on an architecture relying on many remote layers used to anonymize connections and to preserve users’ privacy. Several attacks have been set up against TOR, some complex and often effective. Attacks have diversified over time mostly targeting the network layer, application layer and protocol layer.

Summary

In the paper, Kim et al. present a novel approach of enhancing security and privacy of TOR by using a Trusted Execution Environment (TEE) such as Intel SGX. They make a focus on attacks and information leaks that target TOR components. Indeed, as a volunteer-based network, TOR could allow an attacker to add malicious components such as relays or to compromise such components. The authors consider an adversary as the one being able to modify or extract information from a TOR relay, compromise hardware components such as memory and I/O devices except for the CPU package itself. The authors also consider software components as adversaries, this includes a privileged software such as the operating system, hypervisor and BIOS. They present the security implication of applying a TEE on TOR and evaluate the performance of such method. By adding a new layer a security to TOR components, SGX-TOR prevents code modification and limits the information exposed to untrusted parties at the cost of a performance overhead. Indeed, as presented in the evaluation section, this technique impacts TOR’s performance by increasing the end-to-end latency by 3.9% and the throughput overheads for HTTP connections by 11.9%.

Contribution

By offering anonymity to its users, the TOR network is the target of many attacks. These attacks mainly target its components, especially relays. Indeed, TOR becomes vulnerable when an attacker controls a large fraction of relays. Hence the need to have a large and secure network. In this publication, the authors focus on the security aspect of TOR components by using a TEE, in particular Intel SGX which is increasingly deployed on recent machines with Intel processors. This approach, widely used for sensitive systems (such as location data), adds a layer of security often at the expense of performance, this is the case with the TOR network. The authors apply this novel approach to increase user’s anonymity through the TOR network by thwarting some common attacks described in the summary. Thwarting TOR’s weaknesses to increase its level of privacy is a great way to provide full anonymity online.

Limitation

In this paper, the authors explored the potential offered by Intel SGX in increasing the security of TOR components. However, their solution shows some limitations that can be patched to increase the desirability of the system or to extend the work to other areas of research.

Performance overhead. The technology presented in the paper, although effective in some use cases, remains a nice to have. Indeed, for preserving anonymity, TOR uses relay networks run by volunteers without giving any guarantees of a high level of service. Indeed, one main drawback of the TOR network is its slowness due to its limited capacity and the increasing number of users\(^3\). This matter of fact makes the solution very slow with basic usages. As presented in the evaluation section, by running SGX-TOR in a private environment, the results show that the latency increases by up to 18.4% and the performance reduces by 25.6%. Adding a new layer of complexity affects the performance of the whole system and reduces the interest of the solution.
Conflicting statements. In the scope definition of the paper (Section 3), the authors exclude Denial-of-Service (DoS) attacks and justify it by the ability of the system to simply deny the service (halt or reboot). Meanwhile, in Subsection 3.C, the authors list all the attacks thwarted by SGX-TOR and make a focus on snipper attack. A snipper attack, as presented by Loesing Karsten\(^4\) is a destructive DoS attack that could disable TOR relays by making them to use an arbitrary amount of memory. This attack is based on TOR proxy (client) modification and can be thwarted by SGX-TOR when the client uses SGX.
Compatibility of Intel SGX with TOR components. Another limitation of this paper is that it focuses on an approach where all the TOR main components (clients, directory servers and relays) are using SGX-enabled hardwares. As presented in the paper, SGX-TOR is only compatible with computers running Linux or Windows and having the hardware capabilities to use SGX. This excludes a large number of components, especially TOR Proxies. Regarding other components such as TOR relays, it is worth noting that these components are run by volunteers, TOR does not have a business model and reports a revenue mostly generated by donations and funding from the U.S. Government\(^5\).
Intel SGX as an unbreakable solution. Another limitation of the paper is that the authors present Intel SGX by considering it as being an unbreakable solution to preserve data integrity and privacy. The authors do not question Intel SGX itself. In the literature review, they do not mention nor consider attacks made on Intel SGX. In fact, many attacks such as enclave attacks or spectre-like attacks (Foreshadow attack)\(^6\) have shown the weaknesses of Intel SGX. One interesting attack is a Prime+Probe type attack presented by Schwarz et al. in 2017 that can be used to grab RSA keys from SGX enclaves running on the same system by using CPU instructions in lieu of a fine-grained timer to exploit cache DRAM side-channels\(^7\).

References

Costan, V., & Devadas, S. (2016). Intel SGX Explained. IACR Cryptol. ePrint Arch., 2016(86), 1-118.
Kim, S., Han, J., Ha, J., Kim, T., & Han, D. (2018). SGX-Tor: A Secure and Practical Tor Anonymity Network With SGX Enclaves. IEEE/ACM Transactions on Networking, 26(5), 2174-2187.
http://www.torproject.org/press/presskit/2009-03-11-performance.pdf
Loesing, K. (2009). Privacy-enhancing technologies for private services (Vol. 2). University of Bamberg Press.
https://blog.torproject.org/transparency-openness-and-our-2016-and-2017-financials
Van Bulck, J., Minkin, M., Weisse, O., Genkin, D., Kasikci, B., Piessens, F., … & Strackx, R. (2018). Foreshadow: Extracting the keys to the intel {SGX} kingdom with transient out-of-order execution. In 27th {USENIX} Security Symposium ({USENIX} Security 18) (pp. 991-1008).
Schwarz, M., Weiser, S., Gruss, D., Maurice, C., & Mangard, S. (2017, July). Malware guard extension: Using SGX to conceal cache attacks. In International Conference on Detection of Intrusions and Malware, and Vulnerability Assessment (pp. 3-24). Springer, Cham.