Ontology-Driven Detection of Traffic Light Manipulation in Intelligent Transportation Systems

Intelligent Transportation Systems (ITS) are increasingly centered on Autonomous Vehicles, whose perception stack must be robust against cyberattacks. Recent studies have demonstrated that low-cost laser attacks on traffic lights can bypass recognition systems, resulting in erroneous perceptions of green or red states, thereby endangering road users. Existing Vision-Language Models (VLM) operate on raw visual data and patterns, resulting in the acceptance of visual configurations that, while plausible, are physically or legally unacceptable. To address this challenge, this work introduces an ontology-driven pipeline to verify VLM perception through logical consistency checking. By integrating the VLM with a traffic light ontology developed in this work, the approach establishes a novel pipeline that benefits from both data-driven visual representations and symbolic constraints derived from the knowledge base. This hybrid design can facilitate the identification of inconsistencies in traffic light states.