PHANTOM: Multimodal Adversarial Attacks for Vision-Language Models

PHANTOM is a large-scale, open-source dataset of pre-generated multimodal adversarial attacks. It provides ready-to-use adversarial image–text pairs targeting a wide range of harmful intents for evaluating VLM robustness. The dataset was created by author 'it4lia' and was last updated on Hugging Face in June 2026.

Use Cases

• Benchmarking model robustness based on pre-generated adversarial image–text pairs.
• Systematic safety evaluation of VLMs based on attacks targeting harmful intents.
• Reproducible adversarial testing without generating new attacks.
• Cost-effective safety assessment based on the dataset's ready-to-use nature.

Strengths

• Dataset is described as 'large-scale', suggesting substantial size.
• Provides 'pre-generated' attacks, enabling immediate use for evaluation.
• Targets a 'wide range of harmful intents' for comprehensive safety testing.
• Designed for 'systematic, reproducible, and cost-effective' evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file size are unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the specific adversarial attack generation method.

Provenance

Source

Hugging Face dataset uploaded by author 'it4lia'.

Collection Method

Pre-generated adversarial attacks, likely created algorithmically.

Freshness

Last updated 2026-06-24 08:25:55; freshness should be verified.