MovieTection: 10K-100K Movie Frames for VLM Copyright Detection

MovieTection is a benchmark dataset containing between 10,000 and 100,000 records designed to detect copyrighted visual content in Large Vision-Language Models (VLMs). Developed by the DIS-CO research team in early 2025, it provides image-caption pairs to evaluate model exposure to pretraining data. The dataset supports Membership Inference Attacks (MIA) by testing if models can identify specific movie titles from frames or text descriptions.

Use Cases

• Detecting pretraining data exposure by prompting models to predict movie titles from specific frames
• Evaluating copyright infringement risks in VLMs using the provided image-caption pairs
• Benchmarking Membership Inference Attacks (MIA) on vision-language architectures

Strengths

• Scale of 10,000 to 100,000 records
• Dual-modality support including both image frames and text captions
• Specifically mapped to the DIS-CO research framework for copyright discovery

Limitations

• Geographic bias toward United States cinema as indicated by the US region tag
• Potential for label ambiguity in generic movie frames that lack distinctive visual markers

Provenance

Source

DIS-CO research team, 'DIS-CO: Discovering Copyrighted Content in VLMs Training Data'

Freshness

Last updated March 2025.

Geography

United States