Georgia Institute of Technology
Advancing Automated Program Analysis through the Lens of Program Semantics
Abstract
dc:description.abstractProgram analysis plays a central role in software security. By inspecting code, observing runtime behavior, and reasoning about possible executions, program analysis techniques help developers and security analysts discover bugs, expose vulnerabilities, and improve software reliability. However, applying automated program analysis to complex real-world software remains challenging. Modern programs often process highly structured inputs and execute deep behaviors only after those inputs satisfy layers of syntactic, semantic, and path constraints. Generic analysis techniques, such as grey-box fuzzing and symbolic execution, can scale across many targets, but they often waste effort on invalid inputs, shallow states, or low-value paths. In contrast, techniques that encode target-specific knowledge can reach deeper behaviors, but they require substantial manual effort and are difficult to generalize. This dissertation advances automated program analysis through the lens of program semantics. The central insight is that semantic knowledge, whether explicitly specified by humans or inferred from program context by large language models, can guide automated analysis toward deeper program behaviors. Rather than treating fuzzing and symbolic execution as purely syntactic or path-coverage-driven processes, this dissertation studies how analysis systems can understand what inputs mean, which constraints are worth solving, and which symbolic behaviors should be preserved or pruned during exploration. The first part of this dissertation studies explicit semantic modeling for grey-box fuzzing. We present BuzzBee, a semantics-aware fuzzing framework for database management systems. DBMSs expose diverse interfaces, including SQL queries, key-value command sequences, document operations, and graph queries. Existing generic fuzzers struggle to generate test cases that are not only syntactically valid but also semantically meaningful across these interfaces. BuzzBee addresses this problem by introducing an abstract semantic model based on common data operations, context-sensitive constraint resolution, and dependency-guided mutation. With lightweight human annotations, BuzzBee guides fuzzing toward inputs with stronger semantic meaning and dependencies, and discovers 40 vulnerabilities across 8 DBMSs. The second part studies LLM-inferred semantic guidance for automated program analysis. We present HLPFuzz, a hybrid fuzzing framework for language processors. Language processors impose complex constraints across tokenization, parsing, semantic checking, optimization, and execution stages. HLPFuzz uses an LLM to infer and solve hard constraints that block fuzzing from reaching deep program states, while program-analysis mechanisms prioritize which constraints to solve and iteratively construct concise but sufficient context for the LLM. HLPFuzz discovers 52 bugs across 9 popular language processors and significantly improves code coverage over state-of-the-art fuzzers. We further present Klleem, a symbolic execution framework that uses LLMs to infer program semantics and synthesize assumptions to guide symbolic exploration. Klleem identifies symbolic execution bottlenecks, extracts downstream context, and synthesizes safe assumptions that prune low-value behaviors while preserving behaviors needed for deeper exploration. Together, these three systems demonstrate that semantic reasoning can substantially improve automated program analysis. BuzzBee shows that explicit semantic abstractions can make grey-box fuzzing more effective on structured input domains. HLPFuzz shows that LLMs can infer semantic constraints from source code and use them to generate deep-reaching test inputs. Klleem shows that LLMs can synthesize semantic assumptions to guide symbolic exploration. Overall, this dissertation argues that automated program analysis can become both deeper and more general by integrating semantic knowledge into the exploration loop.
Degree
thesis:*- Name thesis:degree_name
- Computer Science, PhD
- Department thesis:degree_department
- Computer Science
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Yang, Yupeng
- Advisor dc:contributor.advisor
-
- Lee, Wenke
- Committee members dc:contributor.committeemember
-
- Saltaformaggio, Brendan
- Ahamad, Mustaque
- Mertoguno, Sukarno
- Pei, Kexin
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82179