
A Hypervisor Based Security Testbed Dan Duchamp and Greg DeAngelis Computer Science Department Stevens Institute of Technology Hoboken, NJ 07030 USA [email protected], [email protected] ABSTRACT system (IDS) on a real world network that is small and We are developing an experimental testbed intended to help highly vulnerable, such as a honeypot [13] or a DMZ. support security research. The testbed allows a network of Conventional approaches to intrusion detection are unmodified hosts, running any of several of unmodified op- either host-based or network-based. A host-based IDS erating systems, to execute in a controlled and reproducible monitors computing activity on a host and raises an manner. The network is implemented on a hypervisor that is alarm when suspicious activity is observed. One weak- instrumented to observe and control security-relevant events. ness of a host-based IDS is that it places the IDS in These events are securely logged to a relational database for the same domain of execution as the vulnerable ma- later analysis. chine, exposing the IDS to whatever happens when the machine is compromised. Another weakness is that in- 1. INTRODUCTION trusion detection logic is applied to only a single host. To protect an entire installation using host-based IDS We are developing an experimental testbed intended software, each host must run the IDS and anomalous ac- to help support security research. The testbed allows tivity must be detectable on a per-host basis. Attacks a network of unmodified hosts to be subjected to se- that manifest across several hosts may not be detected curity attacks in a controlled, observable, and repro- by a host-based IDS approach. In contrast, a network ducible manner using hypervisor technology. IDS sniffs packets, reconstructs “flows” from the con- The two most notable characteristics of our testbed stituent packets, then applies intrusion detection logic are the use of a virtual machine monitor (VMM, aka to each flow. One weakness of a network-based IDS hypervisor) as a tool for gathering information over an is that flow reconstruction can be very difficult, espe- entire heterogeneous LAN, and a highly generalized in- cially if—as is often the case—an attack intentionally frastructure for the description, measurement, and anal- includes nonsense packets specifically intended to con- ysis of experiments. fuse or trick the recipient. A classic example of an at- Unmodified guest host software (operating system tack containing nonsense packets is when two TCP seg- and applications) runs on an instrumented virtual ma- ments contain overlapping data; e.g., one segment con- chine monitor. The VMM provides a powerful tool tains bytes numbered 100-300, while the next segment for observing and controlling the behavior of the guest contains bytes numbered 200-400. It is implementation- software. Measurement points added to the hypervi- dependent whether TCP will accept bytes 200-300 from sor allow the experimenter to log the occurrence of key the first segment or the second. A network IDS must security-relevant events such as a system call or a buffer first reconstruct the flow that will be seen by the host overflow. When our work is complete, the experimenter under attack, then analyze the flow for anomalies or will be able not only to log but also to control guest host suspicious content. In the case of overlapping TCP behavior via the measurement points; for example, the segments, merely to reconstruct the flow the network- hypervisor could intercept and verify a system call be- based IDS must be aware of the host’s exact TCP state fore allowing it to proceed. and even of the host’s TCP implementation details. The VMM not only emulates x86 hardware, it also Host-based and network-based intrusion detection ap- implements a virtual Ethernet. Emulating both hosts proaches each suffer from their own blind spot. A se- and network on the same hypervisor provides experi- curity testbed facility based on virtual machine (VM) menters with a more complete view of how network- technology need not have a blind spot. based attacks spread. Having a system-wide view permits capture and record- Furthermore, the idea of implementing a LAN and all ing of a timestamped “movie” of activity in all parts of its hosts on a single virtual machine might also prove the system during an attack. Such movies can be an- to be a superior way to deploy an intrusion detection alyzed off-line afterward to discover attack signatures used exclusively to control the physical machines in lim- that could then serve as input to standard IDSs. Movies ited ways such as rebooting them and initializing their could also be stored and replayed later to reproduce an disks. A second “control” network is used exclusively attack on demand. for logging. Events logged by each virtual machine The main components of our testbed are: are transmitted over the control network to a database server (to be described below). 1. Hardware base The third type of network implemented by the switch 2. Infrastructure for experiment description and con- is the “data” network. Guest hosts use only the data figuration network for their traffic. To form networks of guest 3. Support for measurement and logging during exper- hosts, each guest operating system is given a virtual iments network interface. Each virtual network interface is Each component is discussed in its own section below. connected to a software bridge running on the physi- cal machine. The software bridge also connects to its 2. HARDWARE BASE machine’s physical Ethernet adapter for the data net- work. A combination of the settings on the software Our current hardware base is shown in Figure 1. Ex- bridges on each machine and the VLAN settings of the periments run on five real machines, each of which has Catalyst switch allow guest hosts on the same or dif- two dual-core Xeon 5060 (Dempsey) processors with In- ferent physical machines to be grouped arbitrarily into tel Virtualization Technology (IVT) extensions [8, 6] for virtual networks. Traffic for the virtual network passes virtualization and 8GB of RAM. IVT provides VMentry through either one or two hypervisor(s) and therefore and VMexit state transitions that manage where inter- can be monitored.1 rupts and system calls are directed. Certain instruc- Besides the five physical machines that support guest tions must be trapped within a guest host and for- hosts for experiments, two physical machines provide warded to the hypervisor. Access to I/O devices is han- support functions. One machine is used as a database. dled as described below. During experiments, it collects records about security- The virtualization software is KVM [10, 11], a load- relevant events produced by measurement points within able module for Linux versions 2.6.20 and later. KVM KVM and transmitted over the control network. After leverages IVT to implement a virtual machine monitor queueing, these records are stored in a MySQL database. with much less code than traditional hypervisors like The database machine also acts as a disk image server. VMWare [15], Xen [2], etc. KVM expands the tradi- We initialize the disk images of the physical machines tional two modes of a UNIX process (user and system) when a complete physical reset is necessary. to three: user, system, and guest. When guest software The final machine shown in Figure 1 is the control runs in such a 3-state process, I/O instructions execute host. Users connect to this machine to configure the in user mode at the Linux privilege level of the user system in preparation for an experiment, configure and who owns the guest host process. Non-I/O code runs in run experiments, and perform post-experiment analysis guest mode, and system mode is used only to transition of events logged in the database. among modes and for special instructions. The database server and control host are connected Linux/KVM can host any unmodified x86-based op- to only the control and management networks, keep- erating system. Therefore, our testbed supports all op- ing them physically isolated from the dangerous data erating systems of interest to us: Windows, Linux, and network. NetBSD. Support for heterogeneity allows the execution All this hardware is separated from the campus net- of more varied and realistic experiments. work and Internet by a Cisco ASA 5510 firewall. The Because Linux exposes dual dual-core Xeon proces- firewall allows connections between the outside and the sors with hyperthreading enabled as 8 virtual proces- control host so that users can enter the testbed, set up sors, each hypervisor is configured to support 8 guest and run their experiments. hosts by default. So the overall facility supports a min- Experiments running on the data network can also be imum of 40 guest hosts. However, the hypervisors can granted access to the outside when appropriate. Not ev- be configured to support more than 8 guest hosts each, ery experiment will involve dangerous code; the testbed so larger virtual networks are possible. Memory is the can be used for non-security experiments as well as resource most likely to limit the number of virtual hosts for security experiments. The firewall supports 100 that can be supported. Each real machine has 3 Ethernet ports that are all 1 attached to a Cisco Catalyst 3750 switch. The switch Traffic between guest hosts passes through one hypervi- sor if the two guest hosts are running on the same physical implements three types of network. A “management” machine; otherwise, the two guest hosts are running on dif- network connects the 5 real machines to a control server ferent physical machines so their traffic passes through two (to be described below).
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages6 Page
-
File Size-