Image Privacy Risk Benchmark with Ordinal Severity Scores

CPRT-Bench contains approximately 6,700 images annotated for privacy risk assessment. The dataset provides ordinal severity levels and continuous risk scores, sourced from the VISPR (Visual Privacy Advisor) project. It was created by author timtsapras23 and last updated in March 2026.

Use Cases

• Training models to predict ordinal severity levels from image content for privacy classification.
• Benchmarking algorithms that estimate continuous risk scores for fine-grained privacy assessment.
• Evaluating how different visual features contribute to composition-dependent privacy risk predictions.

Strengths

• Approximately 6,700 annotated images provides a substantial benchmark size.
• Includes both ordinal (4 levels) and continuous risk scores for multi-faceted evaluation.

Limitations

• The dataset size of ~6.7K images may be limited for training large-scale deep learning models from scratch.
• Geographic and demographic bias is possible as the source (VISPR) and region tag indicate a US focus.

Provenance

Source

VISPR (Visual Privacy Advisor) project.

Collection Method

Images sourced and annotated from VISPR.

Time Range

null

Freshness

Last updated March 2026.

Geography

Primarily United States (based on platform region tag).