Master Thesis Leveraging Canary Code Techniques for Fact-Checking AI-Generated Software Artifacts in Security-Critical Systems

Problem statement

In systems of high integrity its imperative that Cybersecurity requirements of the system are realized and verified through out the life cycle of the product. We want to explore and understand if the use of Canary Elements can be valid indicators verifying the integrity of the AI (LLM) system used for generating requirements and/or code. Also the in case not using AI for generating artifacts is of interest to be studied.

The systems in scope are embedded systems and not IT systems.

Proposed solution

This Master Thesis investigates the application of canary elements methodologies as a validation mechanism for AI-generated software artifacts specifically targeting code generation and requirements specification in domains with stringent cybersecurity constraints.

The research explores how strategically embedded keywords markers or patterns (canary elements) can serve as verification checkpoints to detect hallucinations ensure compliance with domain-specific constraints and validate the correctness of AI-generated outputs (or any generated output). This approach is particularly crucial in safety-critical systems (automotive aerospace medical devices) and security-sensitive applications where AI-generated code must adhere to strict standards such as ISO 21434 ISO26262 or Common this thesis we focus on the Cybersecurity domain of an embedded system.

Research Objectives

• Conduct a systematic literature review of existing canary code techniques and AI output validation methods
• Analyze current approaches for ensuring reliability in AI-generated software artifacts
• Develop a taxonomy of canary code patterns applicable to different software engineering contexts
• Design and implement a proof-of-concept validation framework using a state-of-the-art Large Language Model
• Evaluate the effectiveness of canary-based validation in detecting AI-generated errors and compliance violations

Outcomes/Deliverables

• Comprehensive survey of existing validation techniques
• Novel canary code methodology tailored for AI-generated software
• Working prototype demonstrating the approach with measurable validation metrics
• Guidelines for implementing canary-based validation in safety/security-critical development workflows

This research contributes to the emerging field of trustworthy AI in software engineering addressing critical gaps in ensuring the reliability of AI-assisted development tools.

You will of course have the opportunity to shape the thesis based on your knowledge skills and discoveries during the project.

Scope of master thesis project

Two students completing 30 credits each (20 weeks) onsite at the Lund office

Qualifications :

Your profile

In order to be successful in the project with think you are:

• Student in Information Technology Computer Science Electronics Math or Physics.
• Required knowledge / courses on data science cybersecurity and AI
• Experienced with or have at least some knowledge of programming in Python C or similar.
• Self-driven able to challenge yourself and gain the experience needed to move the project forward.
• A person with team spirit social skills and a curiosity for exploring new technology areas.

Additional Information :

Why choose Bosch:

In 2022 for the third year in a row Bosch have received the Worlds Best Employer award from Forbes/Statista ranking us among the top 3% of the worlds most attractive employers.

At Bosch we believe that diversity is our strength. We look at diversity in gender generation nationalities and culture as our advantage. We believe mixed teams to be more successful because they utilize the potential offered by different perspectives and solution strategies. We therefore promote mixed teams at all levels and draw on the entire talent pool.

Remote Work :

No

Employment Type :

Part-time