Section: Overall Objectives
Our group focuses on developing automated techniques to compute semantic properties of programs and other systems with a computational semantics in general. Such properties include (but are not limited to) important classes of correctness properties.
Verifying safety critical systems (such as avionics systems) is an important motivation to compute such properties. Indeed, a fault in an avionics system, such as a runtime error in the fly-by-wire command software, may cause an accident, with loss of life. As these systems are also very complex and are developed by large teams and maintained over long periods, their verification has became a crucial challenge. Safety critical systems are not limited to avionics: software runtime errors in cruise control management systems were recently blamed for causing unintended acceleration in certain Toyota models (the case was settled with a 1.2 billion dollars fine in March 2014, after years of investigation and several trials). Similarly, other transportation systems (railway), energy production systems (nuclear power plants, power grid management), and medical systems (pacemakers, surgery and patient monitoring systems) rely on complex software, which should be verified.
Beyond the field of embedded systems, other pieces of software may cause very significant harm in case of bugs, as demonstrated by the Heartbleed security hole: due to a wrong protocol implementation, many websites could leak private information, over years.
An important example of semantic properties is the class of safety properties. A safety property typically specifies that some (undesirable) event will never occur, whatever the execution of the program that is considered. For instance, the absence of runtime error is a very important safety property. Other important classes of semantic properties include liveness properties (i.e., properties that specify that some desirable event will eventually occur) such as termination and security properties, such as the absence of information flows from private to public channels.
All these software semantic properties are not decidable, as can be shown by reduction to the halting problem. Therefore, there is no chance to develop any fully automatic technique able to decide, for any system, whether or not it satisfies some given semantic property.
The classic development techniques used in industry involve testing, which is not sound, as it only gives information about a usually limited test sample: even after successful test-based validation, situations that were untested may generate a problem. Furthermore, testing is costly in the long term, as it should be re-done whenever the system to verify is modified. Machine-assisted verification is another approach which verifies human specified properties. However, this approach also presents a very significant cost, as the annotations required to verify large industrial applications would be huge.
By contrast, the antique group focuses on the design of semantic analysis techniques that should be sound (i.e., compute semantic properties that are satisfied by all executions) and automatic (i.e., with no human interaction), although generally incomplete (i.e., not able to compute the best —in the sense of: most precise— semantic property). As a consequence of incompleteness, we may fail to verify a system that is actually correct. For instance, in the case of verification of absence of runtime error, the analysis may fail to validate a program, which is safe, and emit false alarms (that is reports that possibly dangerous operations were not proved safe), which need to be discharged manually. Even in this case, the analysis provides information about the alarm context, which may help disprove it manually or refine the analysis.
The methods developed by the antique group are not be limited to the analysis of software. We also consider complex biological systems (such as models of signaling pathways, i.e. cascades of protein interactions, which enable signal communication among and within cells), described in higher level languages, and use abstraction techniques to reduce their combinatorial complexity and capture key properties so as to get a better insight in the underlying mechanisms of these systems.