Section: Overall Objectives

Approach and motivation

The claim made by TAMIS is that assessing security requires combining both engineering and formal techniques.

As an example, security exploits may require combining classes of well-known vulnerabilities. The detection of such vulnerabilities can be made via formal approaches, but their successful combination requires human creativity. TAMIS's central goal is thus to demonstrably narrow the gap between the vulnerabilities found using formal verification and the issues found using systems engineering. As a second example, we point out that there are classes of attacks that exploit both the software and hardware parts of a system. Although vulnerabilities can be detected via formal methods in the software part, the impact of attacking the hardware still needs to be modeled. This is often done by observing the effect of parameter changes on the system, and capturing a model of them. To address this situation, the TAMIS team bundled resources from scalable formal verification and secure software engineering for vulnerability analysis, which we extend to provide methods and tools to (a) analyze (binary) code including obfuscated malware, and (b) build secure systems.

Very concrete examples better illustrate the differences and complementarity of engineering and formal techniques. First, it is well-known that formal methods can be used to detect buffer overflows. However, the definition of buffer overflows itself was made first in 1972 when the Computer Security Technology Planning study laid out the technique and claimed that over sizing could be exploited to corrupt a system. This exploit was then popularized in 1988 as one of the exploits used by the Morris worm, and only at that point systematic techniques were developed to detect it. Another example is the work we conducted in attacking smart cards. The very firsts experiments were done at the engineering level, and consisted of retrieving the key of the card in a brute force manner. Based on this knowledge, we generated user test-cases that characterize what should not happen. Later, those were used in a fully automatized model-based testing approach  [39].