Date of Award
Doctor of Philosophy (PhD)
Formal methods techniques, such as verification, analysis, and synthesis,allow programmers to prove properties of their programs, or automatically derive programs from specifications. Making such techniques usable requires care: they must provide useful debugging information, be scalable, and enable automation. This dissertation presents automated analysis and synthesis techniques to ease the debugging of modular verification systems and allow easy access to constraint solvers from functional code. Further, it introduces machine learning based techniques to improve the scalability of off-the-shelf syntax-guided synthesis solvers and techniques to reduce the burden of network administrators writing and analyzing firewalls. We describe the design and implementationof a symbolic execution engine, G2, for non-strict functional languages such as Haskell. We extend G2 to both debug and automate the process of modular verification, and give Haskell programmers easy access to constraints solvers via a library named G2Q. Modular verifiers, such as LiquidHaskell, Dafny, and ESC/Java,allow programmers to write and prove specifications of their code. When a modular verifier fails to verify a program, it is not necessarily because of an actual bug in the program. This is because when verifying a function f, modular verifiers consider only the specification of a called function g, not the actual definition of g. Thus, a modular verifier may fail to prove a true specification of f if the specification of g is too weak. We present a technique, counterfactual symbolic execution, to aid in the debugging of modular verification failures. The approach uses symbolic execution to find concrete counterexamples, in the case of an actual inconsistency between a program and a specification; and abstract counterexamples, in the case that a function specification is too weak. Further, a counterexample-guided inductive synthesis (CEGIS) loop based technique is introduced to fully automate the process of modular verification, by using found counterexamples to automatically infer needed function specifications. The counterfactual symbolic execution and automated specification inference techniques are implemented in G2, and evaluated on existing LiquidHaskell errors and programs. We also leveraged G2 to build a library, G2Q, which allows writing constraint solving problemsdirectly as Haskell code. Users of G2Q can embed specially marked Haskell constraints (Boolean expressions) into their normal Haskell code, while marking some of the variables in the constraint as symbolic. Then, at runtime, G2Q automatically derives values for the symbolic variables that satisfy the constraint, and returns those values to the outside code. Unlike other constraint solving solutions, such as directly calling an SMT solver, G2Q uses symbolic execution to unroll recursive function definitions, and guarantees that the use of G2Q constraints will preserve type correctness. We further consider the problem of synthesizing functions viaa class of tools known as syntax-guided synthesis (SyGuS) solvers. We introduce a machine learning based technique to preprocess SyGuS problems, and reduce the space that the solver must search for a solution in. We demonstrate that the technique speeds up an existing SyGuS solver, CVC4, on a set of SyGuS solver benchmarks. Finally, we describe techniques to ease analysis and repair of firewalls.Firewalls are widely deployed to manage network security. However, firewall systems provide only a primitive interface, in which the specification is given as an ordered list of rules. This makes it hard to manually track and maintain the behavior of a firewall. We introduce a formal semantics for iptables firewall rules via a translation to first-order logic with uninterpreted functions and linear integer arithmetic, which allows encoding of firewalls into a decidable logic. We then describe techniques to automate the analysis and repair of firewalls using SMT solvers, based on user provided specifications of the desired behavior. We evaluate this approach with real world case studies collected from StackOverflow users.
Hallahan, William Triest, "Automated Approaches for Program Verification and Repair" (2022). Yale Graduate School of Arts and Sciences Dissertations. 602.