REVIVE: Multimodal detection and recovery of vandalism in AV

Safety in autonomous vehicles goes beyond avoiding collisions; it must also consider physical attacks such as vandalism on cameras. When a sensor is obstructed with paint, adhesives, or dirt, the perception system loses reliability, which can lead to critical errors. Traditionally, efforts focused on detecting these anomalies, but recovering the useful image remained a challenge. A novel approach combines binary and multiclass detection of the obstruction pattern, semantic segmentation with efficient networks, and three restoration strategies: direct pixel replacement, adaptive median filtering, and a generative inpainting model based on Stable Diffusion guided by linguistic descriptions. This pipeline not only identifies whether an image is vandalized but also applies the most suitable technique according to the type of damage, improving visual quality and, above all, object detection capability. Results show that without recovery, the performance of a YOLOv8 detector drops drastically, while with aligned pixel replacement, accuracy is almost fully recovered. Artificial intelligence for businesses can leverage these developments to protect critical systems in real-world environments. The AI solutions we offer enable building customized visual restoration pipelines and robust decision-making against intentional failures. Furthermore, cybersecurity in these systems covers not only the logical layer but also the physical one: a vandalism attack is a security breach that must be mitigated with automated responses. That is why having cybersecurity and pentesting services helps anticipate vulnerabilities in the perception infrastructure. In this context, companies developing custom software for autonomous vehicles can incorporate these recovery mechanisms as part of their architecture, ensuring that the video stream never reaches the control module degraded. AWS and Azure cloud services facilitate the deployment of these pipelines with high performance, while AI agents can orchestrate asynchronous recovery without blocking real-time perception. Even tools like Power BI can visualize quality metrics of restored images for continuous auditing. At Q2BSTUDIO, we integrate all these capabilities, from custom applications to business intelligence, so that the reliability of autonomous systems does not depend on a single sensor but on a comprehensive resilience strategy.