CVPR 2026@AdvML Workshop Challenge technical report: adversarial multimodal attacks against autonomous driving VLAs

This technical report summarizes the Adversarial Multimodal Attack Challenge held at the CVPR 2026@AdvML Workshop, which targets vision-language agents (VLA) used to interpret driving scenes. The competition is based on DriveLM-style multi-view visual question answering, where each scene consists of six synchronized camera images and a set of structured driving question-answer pairs; participants must generate adversarial images and suffix-only text perturbations that make model answers deviate from the reference answers, while maintaining image fidelity and limiting text cost. The competition runs in two stages, with hidden black-box models added in the second stage to evaluate transferability. The report presents the task design, submission rules, evaluation protocol and leaderboard results, and analyzes the leading solutions from five available technical reports, summarizing several common practices: suffix penalties make image-side attacks more favored; scene-level joint multi-view optimization outperforms handling each view separately; question-answer types and graph structure can provide priors for allocating attack budgets; feature-space objectives can improve black-box transfer; and text content embedded in camera images exposes a persistent weakness of driving VLAs.