RoBERTa 기반 ICS 프로토콜 퍼징 입력 생성 및 제어 기법 = RoBERTa-based controllable input generation for ICS protocol fuzzing

다국어 초록 (Multilingual Abstract)

Industrial Control Systems (ICS) operate critical infrastructures such as power, manufacturing, transportation, and water, and must meet stringent availability and safety requirements. These constraints make it difficult to deploy heavy-weight instrumentation or frequent patching, so simple legacy binary protocols such as Modbus/TCP remain widely used. Automated protocol fuzzing is therefore an important technique for discovering ICS vulnerabilities, but existing approaches often depend on detailed protocol specifications, reduce to near-random mutation that easily breaks message format, or assume source-code access and rich coverage instrumentation that are rarely available in operational ICS. This thesis investigates whether a byte-level language model trained solely on network traffic can serve as a practical input generator for ICS protocol fuzzing. We treat Modbus/TCP as a “language” over byte sequences and train a RoBERTa-based Masked Language Model (MLM). Mask-and-restore generation is then used to produce test inputs, while mask positions, restoration order, sampling temperature, and post-hoc correction of length fields are exposed as simple tuning knobs. Multiple configurations of the proposed fuzzer are compared with existing fuzzers in terms of code coverage, execution efficiency, and request–response behavior. Experimental results show that the RoBERTa-based fuzzer attains code coverage comparable to or better than baseline fuzzers, and that its tuning knobs allow practitioners to balance execution speed, input validity, and diversity without relying on protocol specifications or feedback.

목차 (Table of Contents)

• Ⅰ. 서 론 1
• 1.1. 연구의 필요성
• 1.2. 연구의 목적 및 방법
• 1.3. 논문의 구성
• Ⅱ. 관련 연구 6
• 2.1. 퍼징 기법의 개요와 분류
• 2.2. 프로토콜 퍼징 접근과 한계
• 2.3. 프로토콜의 명세와 언어의 문법
• 2.4. BERT, RoBERTa 개요
• 2.5. ML 기반 퍼징 최근 동향
• Ⅲ. RoBERTa 기반 퍼징 데이터 생성 기법 27
• 3.1. 문제 정의와 설계 개요
• 3.2. 데이터 전처리와 토큰화
• 3.3. 학습 절차: RoBERTa 기반 MLM 학습
• 3.4. 생성 절차: 복원 순서와 온도 제어
• Ⅳ. 실험 및 결과 분석 59
• 4.1. 실험 환경 및 구성
• 4.2. 실험 결과 및 논의
• Ⅴ. 결론 83
• 5.1. 연구 요약 및 의의
• 5.2. 연구의 한계와 향후 연구 과제