SMUGGLEBENCH is a multimodal safety benchmark accompanying the paper 'Making MLLMs Blind: Adversarial Smuggling Attacks in MLLM Content Moderation'. It is designed to study whether Multimodal Large Language Models can identify harmful text hidden, obfuscated, or disguised within images. The dataset was created by author zhihengli-casia and was last updated on Hugging Face in April 2026.
Use Cases
- Benchmarking MLLM robustness against adversarial smuggling attacks based on the described attack pathways.
- Training or fine-tuning content moderation models on visually obfuscated harmful text.
- Researching perceptual blindness in multimodal systems based on the dataset's defined threat model.
- Developing defenses against contextually disguised harmful content in images.
Strengths
- Dataset is explicitly tied to a published research paper, providing academic context.
- Focuses on a defined and organized threat model (Adversarial Smuggling Attacks) with two attack pathways.
- Last updated on 2026-04-08, indicating recent maintenance.
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count, file formats, and license are unknown, which may limit suitability assessment.
Provenance
- Source
- zhihengli-casia on Hugging Face
- Collection Method
- Created to accompany the academic paper 'Making MLLMs Blind: Adversarial Smuggling Attacks in MLLM Content Moderation'.
- Time Range
- null
- Freshness
- Last updated 2026-04-08 12:06:00; freshness should be verified.
- Geography
- null