CS261 Projects

General information

Your term project should address a research issue in computer security and consist of the design of some computer security system or technique, or the analysis and possible improvement of some existing system or technique. The main goal of the project is to do original research on a problem of interest in computer security.

You should work in a small group; I expect that teams of approximately 2--3 will be appropriate for most projects. Of course, expectations will be adjusted according to the number of people in your group. I will not categorically rule out solo teams, but I expect that working in groups will allow you to tackle more substantial research issues. If you have trouble finding a project partner, I can help you get matched up with someone else by maintaining a list of people seeking teammates.

Projects will be evaluated on the quality of their research in computer security. At the end of the semester, you will write a conference-style paper on your work. See below for more details.

I expect that most projects will fall (more or less) into one of two categories:

1. Design. Design projects will usually attempt to solve some interesting problem by proposing a design; implementing a prototype; and using the implementation as a basis for evaluating the proposed system architecture.
2. Analysis. Analysis projects might, for example, study some previously-proposed implementation technique, existing system, or class of systems; evaluate its security properties; find flaws, or strengths, in it; and provide new insight into how to build secure systems.

If you're at a loss for a project topic, I've prepared a list of possible project topics that you can peruse as examples of how to a pick a suitable project. See below. But don't feel limited to these suggestions! They are intended only as examples.

You're welcome to come discuss possible project ideas with me, if you like. I'm happy to make myself available to discuss projects.

A final suggestion: Aim high! The top projects could lead to publication (and in past years, a number of projects have led to publications).

The process

You will write a concise (approximately 1 page) project proposal that should clearly state the problem you will be solving, the key challenges for new research, your plan of attack (including milestones and dates). If there are any special resources you might need from me, mention this as well.

The project proposal was due Tue 15 October.

Here's how to submit your proposal. You should put together a web page for your project; currently all it needs to contain is the project members, title of your project, and proposal. Then just email the URL for your project web page to daw@cs.berkeley.edu by Tuesday, October 15th.

In mid-November I might ask you to write a concise status report so I can make sure the projects are on-track. I am always available to meet with any groups who would like to discuss their project, request additional resources, or ask for advice.

You will also be required to present your project at a poster session, to be held on Monday, December 9th from 2-4pm in the Woz lounge.

Finally, a project report will be due by 9am on Monday, December 16th. No exceptions or extensions will be granted, so get it in on time! See below for instructions.

The final report

You are expected to write a technical paper, in the style of a conference submission, on the research you have done. State the problem you're solving, motivate why it is an important or interesting problem, present your research thoroughly and clearly, compare to any related work that may exist, summarize your research contributions, and draw whatever conclusions may be appropriate. There is no page limit (either minimum or maximum), and reports will be evaluated on technical content (not on length), but I expect most project reports will probably be between 7--15 pages long.

If you are not familiar with writing conference-style papers in computer science (or even if you are), the following resources may help:

• Advice on research and writing
• More resources on writing systems papers
• Detailed advice on technical writing
• Armando Fox's hints on writing good systems papers

You may submit your project report electronically or on paper. I prefer electronic submission, although you may choose either. In either case, the deadline is the same: Monday, December 16, before 9:00am.

If you submit electronically:

1. It must be in a format which is easily readable on Unix platforms: that means HTML, Postscript, or PDF is fine (but not Microsoft Word).
2. Place a link to the file on your project web page (see here for the list of project web pages), and send me email with the URL. I will send you confirmation of receipt.

Example project topics

Note: Some examples are very specific. Others are quite generic; for the generic suggestions, be sure to narrow down the topic substantially and propose something concrete and focused.

If you are interested in any of the project topics below, please talk to me about it; I can make some more concrete suggestions.

New attacks

Find new security weaknesses in any widely-deployed system.

Security auditing

Audit a widely-used and under-scrutinized open-source package that is security-critical. Report on your experiences and lessons. How would you re-structure/re-implement the system to make it more robust? What tools would have made your auditing task easier? How effective are existing tools?

Tools for vulnerability detection

Study ways to build tools to help automate the process of reviewing security-critical source code. Can you use runtime testing, static analysis, model checking, formal verification, or other techniques to detect any interesting classes of common security holes? I can give you some concrete ideas to get you started, if this interests you.

Resilient aggregation

Build "resilient aggregation" components for TinyOS. Sensor networks (such as the TinyOS-based systems studied here at Berkeley) are often used to aggregate sensor information and use the result to control systems. For instance, we might put a sensor in each room of Soda Hall and use the average temperature to control the air conditioning. However, at present, our aggregation operations are not secure against errors or maliciously chosen inputs (as might happen if a sensor is compromised). For instance, consider computing the average of N data values: an adversary who can control one of those inputs can bias the output by any desired amount, hence the "average" is not a resilient aggregation operation. However, the median is resilient, because changing any one input by any amount will cause only a small change in the output. Build a prototype of a few resilient aggregation operators and study their usefulness in some sensor network application.

Adversarial simulation

Build an "adversarial simulator" for TinyOS. Currently, the TOSSIM simulator runs the system under friendly conditions (no dropped packets, no bit errors, no unfriendly timing of the scheduler). Hence, it can't detect ways that an adversary might try to crash or subvert the system. We might be more likely to find security bugs by running the system on parameters that are chosen to be the worst case for system, or that are likely to cover corner cases. For instance, an adversarial simulator might determine ahead of time all possible dependencies between tasks and then test exhaustively all possible orderings of interdependent tasks, looking to see if any of these orderings deadlocks the system. Build a prototype of an adversarial simulator and study its effectiveness.

Preventing casting bugs

Implicit casts are a non-trivial source of security holes in C programs. For instance, the following code is vulnerable:

      typedef unsigned short uid_t;
      void dowork(uid_t u);
      main() {
          int x = read_from_network();
          // Squish root (it's not safe to execute dowork() with uid 0)
          if (x==0)
              exit(1);
          dowork(x);
      }
      

Notice that the check for root is buggy, due to the implicit cast to a 16-bit type in dowork(x): by sending the value 65536 on the network, we can execute dowork(0). There have been vulnerabilities of this form in NFS servers. There have also been vulnerabilities due to overloading of uid -1 (on some systems, uid 65535 was the nobody account, but the set*uid() calls treated -1 as meaning "don't change this uid"). There have also been many security holes due to implicit casts between signed and unsigned types. Is there anything intelligent and cost-effective we can do about any of these risks?

Validation bugs in Linux kernel

Dawson Engler's group found many security holes in the Linux kernel based on improper input validation (see their paper) based on heuristic compile-time rules. Formalize their rules into a concrete type system, perhaps using flow-sensitive type qualifiers. Then, analyze the Linux source code using your type system, perhaps using Cqual for the type-checking.

Inlined reference monitors

Proof-carrying code takes a reference monitor (expressing some security policy) and injects it into the program during compilation, in an integrated way; then it is possible for the recipient to verify that the desired policy is enforced by the presented code. This allows us to build extremely efficient reference monitors: rather than implementing the reference monitor in a separate process (thereby incurring performance penalties), the reference monitor can be inlined directly into the code that it is supposed to monitor. Sometimes we don't care about the ability of a recipient to verify that code injection was done appropriately, for instance because the injector and receiver are one and the same. Can we build a more lightweight implementation of this functionality?
For instance, maybe we can emulate some notion of "user-level call gates" using the mprotect() system call? (We could imagine storing the reference monitor's state in a write-protected segment of memory; then any call to an interface that is protected by the reference monitor would be replaced with a call to the reference monitor entry point; and the entry point would need to write-enable the special segment and atomically transfer control to the reference monitor code.)

Virtual machines for security

Recently, software has become available to implement a virtual machine for modern operating systems (e.g., Windows). This seems to provide a powerful mechanism for executing dangerous actions in an isolated environment. Does this idea work, and if so, how can we best take advantage of virtual machine techniques? Can we evaluate the security of, say, the VMWare virtual machine against malicious attempst to harm the host OS? Is there any way to structure the the virtual machine implementation to isolate the security-critical functionality and thereby make the TCB simpler and easier-to-verify?

Enforcing resource bounds on malicious code

Can we use proof-carrying code techniques to ensure that malicious code never exceeds a fixed resource bound? For instance, we might insist that it terminate within a given number of clock cycles; we could insert checks to a global timer whereever we cannot prove a satisfactory upper bound on the running time of the program, and omit the checks in regions (e.g., acyclic control-flow graphs) where we can verify statically that the time bound will not be exceeded.

Side channel attacks meet mobile code

Typically the easiest way to break a cryptosystem is not by directly attacking the mathematics but by "cheating" (e.g., bypassing the crypto entirely), and one way to do this is to use side channel attacks. Java and other forms of mobile code give attackers a convenient way to run code on targetted machines, which might give the attacker a way to measure timings or memory operations by observing scheduling or swapping decisions. The goal of the project would be to investigate whether it is possible to mount side channel attacks within the constraints imposed by Java or some other widely-deployed mobile code system.

Verifiable distributed computation

The Internet is a vast resource of idle machines; we might like to harness these spare CPU cycles by offloading our lengthy computations to other computers. But in any such distributed setting, how do we know that the result that comes back is the correct one we wanted? Careful engineering combined with some recently-proposed cryptographic techniques might go a long way here in solving some cases of interest.

Security of peer-to-peer systems

Peer-to-peer systems (e.g., Gnutella, Kazaa) have been a hot topic recently. You might study the security challenges inherent in peer-to-peer systems, either by proposing techniques for building secure peer-to-peer systems, or by analyzing an existing peer-to-peer system.

Information retrieval for audit logs

Suppose you have audit logs of, say, network events. How would you design a search engine so you could retrieve security-relevant events after the fact? What network events would you want to have logged, and in what format? For instance, imagine tomorrow CERT announces a new attack, and CS admins discover that one of their machines has been broken into. It would be nice if we could just search all audit logs over the past year to see if any other CS machines had been broken into using the same attack. What can you do in this area?

Formal modelling of security systems

Build a formal model of some aspect of a security system, and rigorously evaluate its properties. For instance, you might look at the state machine associated with a TCP/IP stack, and model how the various network events can affect the state. You could build a formal model of actual behavior by working from the OS source code or by exhaustively testing the possibilities. Then you might build a formal model of intended behavior -- e.g., by working from the RFC, or by formalizing that there should be no LAND attacks (for instance) -- and you would check whether the specification matches the verification. Or, you might build a second model from a second operating system, compare where their behavior differs, and study whether this has any consequences for how to write portable security code. Such models might also be useful for intrusion detection as well.

Secure coding

Improve (somehow) the state of the art in implementation of security-critical projects. You might explore the relevance of various techniques from software engineering or programming languages, for instance.

Theory or cryptographic work

Projects in theory or cryptography are not necessarily out of scope, if you have some specific ideas you consider relevant. For instance, formal analysis of cryptographic protocols or cryptographic primitives could probably serve as a reasonable topic.

Privilege separation

We saw in class a number of mechanisms for building a sandbox to prevent untrusted code from affecting the rest of the system in any way. This is a useful primitive, but in practice we often want to allow some amount of controlled sharing or limited interaction. You might study how to securely allow this limited interaction in some application context of interest, such as dividing programs into privileged and unprivileged pieces. The LSM project for Linux might be of interest here.

Distributed firewalls

Imagine taking the enforcement mechanism found on a firewall and replicating it on all inside machines. (This would require installation of new software on all internal machines, but suppose we can handle that administrative burden.) How do we maintain centralized control over security policy? How do we specify policies? How do we maximize assurance in such an environment? How do we handle multi-organization scenarios, where a machine is a member of multiple organizations and thus multiple parties would like to add security restrictions?

Encrypted databases

Suppose we want to store our data on a remote server (e.g., so that we can take advantage of the computational power of the server) without requiring full trust in the remote server. What types of database semantics can we support? In previous work, Dawn Song, Adrian Perrig, and I showed a special way to store the data in encrypted form on the untrusted server so that the trusted server can do keyword searches over the encrypted data (on our behalf, and only when authorized by us) with minimal communication complexity. What other types of queries can be supported efficiently? How about if the goal is integrity rather than confidentiality? This is a cryptographic design problem.