Download Understanding System Security: The Role of Machine Learning - Prof. David Primeaux and more Study notes Computer Science in PDF only on Docsity! Primeaux’ Musings A system is associated with a set of desired behaviors. A risk to a system represents the potential loss of some desired system behavior. A cost is often associated with a risk. A threat to a system is a specific exposure to a specific risk. A system is secure to the extent that threats to the system are reduced. Some systems are more secure than others. However, complete elimination of threat is not possible; therefore, an absolutely secure system is not attainable. So, when people talk about a secure system they are talking about one of three things: An impossibility An ideal that, while not attainable, provides an understandable goal A system that claims to adequately addresses a specific threat Some maxims to think about: “A ship is safe in harbor, but that is not what ships are for.” By isolating a system on a deserted island with no humans around, we might create a nearly perfectly secure system [don’t forget the possibility of earthquake, tsunami, and lightning!], but such a system would not exhibit all (or even many) desired behaviors. When we put a system to use, that system is threatened. “The architecture of system security should be structured in layers, like an onion.” This means that, just as an onion presents itself as a layered structure, layer after layer of security measures are incorporated in a system. Another way of describing this approach is that, to address a specific risk, we should have in place a “chain” or sequence of measures. As a simple example, we might have user authentication by username, followed by password, followed by a biometric measure, followed by scanning of some RFD device. If anyone of these links in the chain “breaks,” the user is not authenticated. From another perspective, “A chain is only as strong as its weakest link.” Suppose a system has no network connections. Then this might be seen as a robust link is a chain. However, often the weakest link in system security is an authorized human user who either unintentionally or maliciously behaves in such a way as to incur cost by reducing desired system behavior. Note, too, that humans are subject to “social engineering” attacks. Further enhancing other “links” in this chain of security without addressing the weaker links might not be cost effective. “Security has associated costs.” When we look at the cost of a security measure, we should include not only the cost of implementation and maintenance, but also the extent to which the security measure itself adversely affects the system’s desired behavior. The question to address in this course is: How can we use machine learning techniques to enhance a system’s security? Or, in terms that we can measure: How can we use machine learning techniques to reduce threats (and their associated costs) to a system? Below is a matrix relating some aspects of systems to security issues that can be addressed by machine learning: Related to Issue User Data Processing, transport, access Protection X X Integrity X X Privacy X X Confidentiality X X X Availability X X X Non- repudiation X Authentication X X Assurance X X Verification X X Intrusion Detection X Authorization X X Permission X X X Recovery X X Student presentations: (Notes: Consider your approach to your topic in the context of other topics. Thus, if you select “Support vector machines to detect malicious executables,” you should not include much, if any, information you would expect another student to have presented in the topic “What is a malicious executable?” Remember: our goal is to have non-overlapping presentations. Also, the questions listed with each topic are mere suggestions and should certainly not be viewed as exhaustive! And, as you look at a topic, creatively ask whether the approach might be “transferred” to other issues or components of a system. Each topic must be approached in the context of machine learning! For each presentation, indicate open questions that might benefit from additional exploration.