The use of virtualization by cloud service providers to host virtual machines belonging to multiple customers on a shared physical infrastructure is opening up fresh data leak risks, a research report warns.
The report by four researchers at MIT and the University of California at San Diego shows how vulnerabilities in cloud infrastructures could allow attackers to locate and eavesdrop on targeted virtual machines (VMs) anywhere in the cloud.
The attack described in the report was conducted against Amazon's Elastic Computer Cloud (EC2) service. But the vulnerabilities that enable it are generic and would likely affect other cloud providers, said Eran Tromer, a post-doctoral researcher at MIT's Computer Science and Artificial Intelligence Laboratory and one of the authors of the report. The report is scheduled to be presented at the Association for Computing Machinery (ACM) Conference on Computer and Communications Security next month.
The research raises questions about a fundamental assumption about cloud computing which says that data hosted in a cloud is relatively safe from targeted attacks because it's hard to know where in the cloud the data is located. The reserach also comes at a time when concerns are high about security and privacy issues related to cloud computing.
According to Tromer, the research shows that it is possible for attackers to identify the physical server on which a targeted virtual machine is hosted in the cloud. The attackers can then establish a rogue virtual machine on the same machine to go after the victim. A virtual machine is an operating environment created within another larger environment. A VM acts as a self-contained computer within a larger server, with virtual boundaries separating each VM from the other. Multiple VMs can run within one physical server.
The multi-stage attack starts with mapping the internal cloud infrastructure to locate the physical server hosting a target VM. Much of the information needed to glean the location of a target VM hosted in a cloud is contained in the IP address and domain name for that particular machine, Tromer said.
In the case of Amazon's EC2 infrastructure, for instance, analyzing the IP address of a VM can reveal details such as geographic region, as well as the availability zones or specific infrastructure segment it is on, he said.
The IP address also specifies an instance type, indicating the amount of computational power, memory and persistent storage that is available to the virtual machine. In addition, VMs located on the same physical server also tend to have IP addresses that are close to each other and are assigned at the same time.
The data gives attackers an idea of the parameters needed to establish a rogue VM on the same physical server as the target VM. They can then proceed to do this by instantiating new VMs until one is placed "co-resident with the target server," Tromer said.
Attackers can significantly boost their chances of achieving "co-residency" by launching a denial-of service-attack against the target server and forcing it to expand capacity by adding new VMs. If the hackers simultaneously request new VMs of their own, their chances of getting one on the same physical machine as the target, is significantly increased.
According to Tromer, once an attacker gains access to the same physical server as the target VM, the attacker can monitor shared resources on the server to make highly educated inferences about the target VM.
For instance, by monitoring CPU and memory cache utilization on the shared server, an attacker could determine periods of high activity on the target servers, estimate high-traffic rates and even launch keystroke timing attacks to gather passwords and other data from the target server, Tromer said. These "side-channel attacks" have proved highly successful in non-cloud contexts so there's no reason why they shouldn't work in a cloud environment, he said.
"The basic vulnerabilities, such as architectural side-channels, are inherent to virtualization technology used by all infrastructure-as-a-service cloud providers," Tromer said.
What the research shows is that until cloud providers can guarantee impermeable partitions between virtual machines on a single server, customers should try as much as possible to avoid sharing physical servers with others in the cloud, he added.
Amazon did not respond to requests for comment. But in comments made to the MIT Technology Review, a spokesman said that Amazon has already rolled out safeguards to protect against the mapping techniques described in the research paper.
The company also refuted the notion that side-channel methods could be used to steal information from a VM on a shared physical server. In comments to the MIT Review, the Amazon spokesman said the researchers had tested such attacks in a "carefully controlled lab configuration that do not match the Amazon EC2 environment."