PHANTOM Dataset Released for VLM Adversarial Attack Research

The research community now has access to PHANTOM, a new large-scale, open-source dataset specifically designed for multimodal adversarial attacks targeting vision-language models (VLMs). This dataset addresses the significant computational cost and complexity typically associated with generating a large volume of adversarial samples. PHANTOM is diverse, representative, and practical, extending existing benchmarks by encompassing 10 high-level categories and 55 subcategories of harmful intents. Comprising 47,524 adversarial samples, the dataset was generated using state-of-the-art attack strategies from recent literature. It consolidates and expands upon prior benchmarks, offering 7,826 distinct intents and introducing an additional category to broaden coverage. The primary goal of PHANTOM is to provide realistic evaluation resources for researchers and practitioners to systematically assess VLM robustness and safety, fine-tune attack-generation models, and develop or stress-test defensive guardrails under a wide array of adversarial conditions.