16.5 Coming Trends
So far, we have looked at the basic foundations of physical security and some of the ways it is embodied in tools available for system design. We close the chapter by looking at some new trends.
16.5.1 Virtualization and Security
Much security research—both old and new—is driven by a basic challenge. It can be useful to have separate compartments within a machine, with high assurance that malicious code in one compartment cannot spy on or disrupt the others. However, we often don’t want complete separation between the compartments—but we want to make sure that only the right types of interaction occur. How do we provide this separation? How do we mediate the interaction? How do we provide assurance that this all works, that this all provides the desired properties, and that it doesn’t kill performance?
In some sense, the challenge motivated the notion of separate processes within an operating system. However, most of the field has accepted the unfortunate notion that the standard OS model will not provide an appropriate solution here. This conclusion comes from several beliefs. Standard operating systems provide too rich and complex an interaction space between processes; standard operating systems are written too carelessly; target applications require more than simply the OS-level interface.
As we discussed in Section 16.4.2, these drawbacks have led to renewed thinking about virtualization: other ways to provide these separate virtual compartments and to mediate interaction between them. However, it’s not clear what the "right" way to do this is. Right now, in both academia and industry, we see lots of approaches swirling around to enable machines to have highly compartmented pieces. This exploration raises many issues.
- What’s the right level to split the machine into compartments?
- Should we use hardware-based support?
- Should we use a virtual machine monitor/hypervisor running above the hardware? (This is called a type I virtual machine—recall Figure 16.4.)
- Should we use a virtual machine monitor running within or above the host OS? (This is called a type II virtual machine—see Figure 16.5.)
Figure 16.5 In the type II approach to virtualization, the VMM runs above the host operating system.
- Should we instead virtualize some image above the OS? (Examples here include UML, BSD Jails, and Solaris Zones—see Figure 16.6.) We have seen some researchers call this approach paenevirtualization.
Figure 16.6 In yet another approach to virtualization, an enhanced OS groups processes into sets called zones or containers (above). One OS installation manages all the containers—however, from the perspective of the userland processes, each container appears to have its own instance of the OS and the machine (below).
- Does the guest software—in particular, the OS—need to be rewritten in order to accommodate the virtualization, or it can be run unmodified? Paravirtualization refers to the former approach.
Some projects to watch in this space include VMWare [VMW07] and XEN [BDF+03, Xen07].
Another set of issues arise pertaining to the mediation between the compartments. How do we define the APIs? How do we have assurance that the APIs, if they work as advertised, work as intended? How do we have assurance that they are implemented correctly? Furthermore, an often neglected issue is how easily human designers and programmers can craft policies that capture the intended behavior. One might remember that these same issues vexed OS design—and one might cynically ask why virtualization research will do any better.
As we discussed earlier, one can take many approaches to providing this illusion. However, if the goal is to provide the illusion of the conventional architecture, doing so with conventional architecture is problematic. The guest system running inside a virtual machine expects to have user and kernel privileges. If guest kernel mode runs inside real user mode, then kernel-mode instructions won’t necessarily operate properly. But if guest kernel mode runs inside real kernel mode, then nothing stops the guest from interfering with other virtual machines. (Indeed, the hacker community celebrates its red pills:6 techniques to determine whether a program is running in a virtualized environment, via exploiting such differences. These properties were well known in the security literature, however.) For that matter, even if we could straighten out the privileges for virtual machines, how do we manage them all? Which code should do that, and how does the architecture enforce that? ([U+05] provides a nice overview of these challenges.)
Hardware vendors have been developing a new generation of processor architectures to address these issues. Intel’s Vanderpol technology (VT), sometimes defined as virtualization technology, removes the privilege and address space obstacles to virtualization. Intel’s LaGrande technology (LT) adds support for secure management of virtualization; essentially, turning the kernel/user model into a quadrant: kernel/user for VM and kernel/user for hypervisor, which has special privileges. As of this writing, LT details are still limited to innuendo; according to rumor, the reason VT and LT seem to overlap somewhat in functionality is that they were rival, not complementary, projects. (Also, new marketing terms, such as TXT, appear to be supplanting these older project names.) AMD’s Pacifica and Presidio architectures correspond to VT and LT, respectively.
Although not necessarily designed for security, these virtualization-friendly architectures have security applications. Such platforms as an IE/Windows Web-browsing installation, which typically are lost causes for security, can be safely confined in a VM. Watchdog modules that check system integrity no longer have to be constrained to looking at RAM and guessing at the internal state of the CPU; they can run inside the CPU, with their fingers inside the target VM. (Virtual machine introspection is a term used for this sort of thinking.) On the other hand, the SubVirt project illustrates another type of security application: malware that inserts itself as a malicious hypervisor (wags suggested calling it a "hypervirus") and shifts the victim system into a virtual machine, where it can’t easily detect or counteract the malware [KCW+06].
16.5.2 Attestation and Authentication
We began our discussion of hardware-based security by considering how hardware can bind secrets to the correct computational entity. We see the potential for much industrial churn and ongoing research here.
One of the first issues to consider is which a-word applies: attestation or authentication or perhaps even authorization. When considering the general scenario of interacting with a remote party, the primary security question is: Who is it? Is it the party I think it is? Resolving this question is generally considered the domain of authentication. As discussed in Section 9.7.2, some dissidents instead see authentication as addressing the binding between entity and name and preach that authorization, as the binding of entity to property, is the true goal.
When the entity in question is a computational device, both identity and properties depend in part on the basic configuration of the device: what the hardware is and what the software is. However, as some of our own work demonstrated (e.g., [Smi04a]), the question is more subtle. The "correct" device can and perhaps should change software while remaining the same entity; the same hardware with a fresh reinstallation of the same software may in fact be a different entity. Consequently, consideration of the a-words in this realm often leads to an initial focus on attesting to a manifest of software and hardware configuration.
Some colleagues insist that, if checking software configuration is involved, then it must be attestation. We disagree.
- One cannot have meaningful attestation without authentication. An entity can claim any configuration it wants to; without authentication, the relying party has no reason to believe this claim.
- On the other hand, one can easily have authentication without attestation. When connecting to a hardware-hardened Web server, what the relying party cares about is the fact that it’s a hardened Web server. The relying party does not necessarily need to know the full state of libraries and Apache versions—just that they’re okay.
Whichever a-word one uses, we see two current sets of unresolved problems. The first is the "right" way to implement this structure within a multicompartmented machine. (We use the general term multicompartmented because we see this applying to a range of beasts, from SELinux boxes with TPMs to advanced virtualization and multicore.) Should a hardware-based root provide a full manifest for each compartment? Should a hardware-based root provide a manifest for a software-based root that in turn certifies each compartment? (And for that matter, why not other combinations of one or more hardware roots with one or more software roots?)
The second set of unresolved problems pertains to what should get listed in a manifest. What is it that the relying party really wants to know about a remote machine? We often joke that giving a TCG-style set of hashes is akin to the uniformed person at the door providing a DNA sample when asked to prove that he or she is really a bona fide police officer—it’s a detailed answer that does not really give the right information. Current researchers are exploring property-based attestation, based on third parties’ providing bindings, and semantic remote attestation, based on programming language semantics. This space will be interesting.
16.5.3 The Future of Moore’s Law
In 1965, Gordon Moore observed that the number of tranistors on an integrated circuit doubles every 2 years. Subsequently blurred and transformed (e.g., the time-line is often presented as every 18 months), this curve has now entered popular folklore as Moore’s Law,7 usually stated as "every N years, the number of transistors on chips will double."
So far, industry has stayed true to Moore’s Law. However, insiders (e.g., [Col05]) observe that the causality is a bit more complicated than might meet the eye. Yes, Moore’s Law was a good predictor of the trend of technology. But also, the industry came to use Moore’s Law as a road map for its business model. For example, the generation N processor might currently be manufactured and the generation N +1 almost ready to fab; however, the design for the generation N + 2 processor, to be fabbed k years later, was already under way and was counting on the fact that chip technology supporting far more transistors would be ready when the processor was ready to be manufactured.
Recently, hardware researchers have begun to express concern about the future of Moore’s Law. Among many, the conventional wisdom is that, in order for Moore’s Law to continue to hold, the transistors themselves will become less reliable—in terms of increased failure rate during manufacture and also, perhaps, in terms of increased failure rate in the field.
Some conjecture that the increased failure rate at manufacture will lead to a stronger emphasis on multicore devices. Committing to one large monolithic processor is risky, since faulty transistors might render the entire chip useless. An architecture that instead consisted of many smaller, somewhat independent modules is safer—the vendor can include a few extra modules in the chip, and sufficiently many should turn out to be good, even with faulty transistors.
However, we might also conjecture that an increased failure rate in the field might lead to a resurgance of work on Bellcore attacks and countermeasures (recall the discussion in Section 16.2.1).
16.5.4 Personal Tokens of the Future
The personal tokens common in the past decade were smart cards: credit-card-sized pieces of plastic with small chips on them, typically used in small-value commercial transactions. As we observed in Section 16.3.4, USB devices are common now. What’s coming next?
Personal digital assistants (PDAs) are one possible candidate. For designers of security protocols, PDAs offer the advantage of having an I/O channel the user trusts, thus avoiding some of the problems of traditional smart cards. However, one might be cynical as well. As PDAs become more like general-purpose computing environments, the greater their risk of contamination—and the less advantage they offer over a risky general-purpose platform. Some economic observers predict that cell phones will displace PDAs. For the security designer, cell phones offer the challenge that it can be harder to experiment and deploy new applications; vendors tend to keep things locked up. (Looking at the students and young professionals who surround us, we might wonder whether iPods might be usable as a personal token.)
The burgeoning use of RFID (radio frequency identification) devices also offers potential for security applications and abuses. Of course, the first step in this discussion is to nail down exactly what RFID devices are. Everyone agrees that these are electronic devices that use some type of close-range radio to communicate. However, the sophistication assigned to these devices varies, from simple replacements for optical barcodes to more complex devices that are armed with environmental sensors, state, and batteries and that participate in interactive protocols with "reader" devices.
Anyone who has tried to wrestle groceries into the right position for a laser to read the barcode printed on them can immediately appreciate the advantages of inexpensive RFID tags that can be read from any orientation, without line-of-sight. Indeed, discussions of application scenarios often begin on such use cases: replacing clumsy optically read tags with easy and efficient RF-read ones, on items such as groceries, library books, warehouse inventory, and passports. No need to manually wrestle the item into the right position—the RF makes the connection automatically!
Of course, the same ease of use that motivates the application of RFID technology is also the source of its security and privacy worries. A machine-readable barcode is typically big enough for a human to see as well—so humans can be aware of its presence. The physical manipulation required for a barcode to be scanned is typically big enough for a human to notice—so humans can make judgments about what’s being scanned and when. Humans also understand the notion of "sight" and thus have a good intuition of how to keep an optical tag from being seen.
These artifacts, which made it possible for human end users to control and understand the use of optical identifiers, disappear with RFID. Which objects have tags? Who is reading them and when and from how far away? What are the privacy implications of this quantum leap in automated information gathering?
Of course, the general notion of an inexpensive device communicating over an open medium raises the more standard security questions of physical security of the end device and communications security between them.
Juels’s survey [Jue06] and the Garfinkel-Rosenberg anthology [GR05] provide more discussion of this problem space. Recently, NIST even published guidelines for RFID security [KEB+07].