LLM-based Property-based Test Generation for Guardrailing Cyber-Physical Systems

Abstract: Cyber-physical systems (CPSs) are complex systems that integrate physical, computational, and communication subsystems. The heterogeneous nature of these systems makes their safety assurance challenging. In this paper, we propose a novel automated approach for guardrailing cyber-physical systems using property-based tests (PBTs) generated by LLMs. Our approach employs an LLM to extract properties from the code and documentation of CPSs. Next, we use the LLM to generate PBTs that verify the extracted properties on the CPS. The generated PBTs have two uses. First, they are used to test the CPS before it is deployed, i.e., at design time. Secondly, these PBTs can be used after deployment, i.e., at run time, to monitor the behavior of the system and guardrail it against unsafe states. We implement our approach in ChekProp and conduct preliminary experiments to evaluate the generated PBTs in terms of their relevance (how well they match manually crafted properties), executability (how many run with minimal manual modification), and effectiveness (coverage of the input space partitions). The results of our experiments and evaluation demonstrate a promising path forward for creating guardrails for CPSs using LLM-generated property-based tests.

An Exploratory Study of LLM-Based Property-Based Test Generation for Guardrailing Cyber-Physical Systems

The paper "LLM-based Property-based Test Generation for Guardrailing Cyber-Physical Systems" by Etemadi et al. explores an automated approach to generating property-based tests (PBTs) for cyber-physical systems (CPSs) using LLMs. The need to ensure safety and robustness in CPS stems from their intricate integration of computational and physical components, which makes their testing and validation particularly challenging. Instead of conventional techniques, this paper investigates employing LLMs to synthesize PBTs that help ensure a CPS operates within safe parameters, effectively guardrailing systems against potential unsafe states.

Approach and Implementation

The methodology proposed in the paper involves a two-phase process:

1. PBT Generation Phase: Leveraging the strengths of LLMs in understanding complex code and documentation, the system ChekProp generates PBTs using LLMs. ChekProp synthesizes initial prompts from the source code and accompanying documentation to guide the LLM in creating PBTs. It incorporates generated unit tests to refine and improve the generated PBTs iteratively, addressing any identified shortcomings.
2. Property-Based Monitoring Phase: Once the PBTs are generated, they serve as runtime assurances, continuously verifying that the CPS remains within safe operational bounds during actual deployment.

The ChekProp tool stands at the core of this process. Written for Python-based CPS applications, ChekProp utilizes Gemini LLM to read and interpret requirements, transforming them into effective tests. Emphasis is placed on using hypothesis-based property tests which offer robust verification across potential runtime scenarios against predefined property requirements, both at the design phase and during actual execution.

Findings and Implications

The efficacy of ChekProp is measured against CPS applications from three different domains. The assessment centers on the relevance of the properties extracted automatically by the LLM and the quality of the generated PBTs:

• Relevance of Extracted Properties: The success lies in LLM’s ability to discern critical properties often manually crafted by humans. ChekProp was able to extract 94% of manually defined properties across the test cases, showcasing promise for automating an otherwise manual process.
• Quality of Generated PBTs: Similarly noteworthy, 47% of the generated tests were executable without major revisions, and 85% effectively covered input space partitions, underscoring their robust nature in verifying CPSs at runtime.

These results demonstrate that LLMs possess significant potential in reducing human effort associated with generating test specifications, especially for CPSs where safety and efficacy hold paramount importance.

Challenges and Future Prospects

Challenges encountered include the need for accurate environmental mocking and nuanced interactions within CPS systems, areas where LLM-generated outputs may sometimes falter due to insufficient contextual information. As a response, the paper advocates for improved prompt engineering, potentially through few-shot learning examples, and comprehensive documentation in model usage to further enhance LLM output accuracy.

Looking forward, these findings pave the way for a paradigm shift in automated software verification—specifically, combining AI-driven model understanding with traditional testing frameworks to augment CPS safety. With the continual evolution of LLM capabilities, future explorations might include more intricate environments, potentially extending beyond controlling unsafe states to predicting emerging risks within complex systems.

In summary, this paper initiates a promising avenue for AI applications amidst rigorous physical and computational settings and holds significance for advancing automated safety mechanisms within the domain of CPS.