ProvDent: 58,320 Inference Call Records for Prompt Injection Attacks on Dental VLMs

58,320 structured JSON records from a study of image embedded prompt injection vulnerability and defense effectiveness across four vision-language models applied to dental panoramic radiography. The dataset includes 9,720 baseline calls and 48,600 defense calls, with pre-computed analysis tables. It was authored by Babak Saravi and last updated on April 10, 2026.

Use Cases

• Benchmarking defense mechanisms against prompt injection attacks based on the 48,600 defense call records.
• Analyzing model vulnerability to adversarial image prompts based on the comparison of baseline and attack calls.
• Studying the behavior of vision-language models on dental panoramic radiography under adversarial conditions.

Strengths

• 58,320 structured inference call records provide a substantial basis for analysis.
• Includes a clear breakdown of 9,720 baseline calls and 48,600 defense calls for controlled comparison.
• Pre-computed analysis tables are included, which may reduce initial processing effort.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual tables is unknown, which may limit suitability assessment.
• The dataset's 6.3 MB size suggests it contains structured records rather than the original image data.

Provenance

Source

Babak Saravi via figshare.

Collection Method

Inference call records generated from a study of image-embedded prompt injection on four vision-language models.

Time Range

Study period not specified; dataset reflects a specific experimental timeframe.

Freshness

Last updated 2026-04-10 22:36:22; freshness should be verified.

Geography

Spatial coverage is not specified; the data is likely model-focused rather than geographic.