matybohacek/t2v-mia

VideoUFO-NonMember

Data release for Membership Inference for Generative Video Models by Raina Panda and Maty Bohacek. Code: github.com/maty-bohacek/t2v-mia.

VideoUFO-NonMember is a set of non-members complementing VideoUFO, constructed by re-running the VideoUFO collection pipeline on Creative Commons YouTube videos uploaded after June 1, 2025. None of its source videos appear among the 345,437 source videos of VideoUFO. Members and non-members are captioned with the same model (Qwen2.5-VL-7B-Instruct) and paired such that blind classifiers, which never query the model, can barely tell them apart. The release also contains the per-clip outputs of every attack in the paper on the 500 evaluation pairs, which suffice to reproduce our tables on a CPU.

Since this data allows one to probe whether specific videos of real people were used to train a model, access is reviewed by the research team.

Contents

Path	Contents
nonmember_clips/	All 18,387 non-member clips from 10,007 source videos across 1,286 VideoUFO topics, named <YouTube id>.<k>.mp4
metadata/nonmember_collection.jsonl	One record per attempted source video: YouTube id, topic, upload date, duration, and the start and duration of each kept clip
pairs/matched_pairs.json	The 4,222 matched member/non-member pairs and the matching setting
pairs/pairs.json, pairs/pairs.csv	The 500 evaluation pairs, [non-member, member, match cost], with topics and captions in the CSV
pairs/captions.csv	Qwen2.5-VL-7B-Instruct captions of the 1,000 evaluation clips
videos/members/, videos/nonmembers/	The 1,000 evaluation clips. Members are original VideoUFO clips, bit-identical to the training data
metadata/split_clip_metadata.csv	VideoUFO metadata of the 500 evaluation members
metadata/topic_frequency.csv	Number of VideoUFO training clips per topic
blind/	Blind features: encoding and quality statistics, CLIP and DINOv2 embeddings for the whole member pool (21,940 clips) and all collected non-members, and SigLIP2, color, spectral, and temporal features for the evaluation clips
activations/	Representation read-out features of the target (acts_target.npz), the reference model, the randomly initialized model, and the target on re-encoded clips; per-block gradient norms and loss (wb_target.npz)
losses/	Forward and time-reversed loss surfaces of the target and the reference model; prefix-completion errors
scores/	Per-clip read-out scores

The layout matches the --data argument of the evaluation scripts in the code repository.

Licensing

The non-member clips are excerpts of YouTube videos published under the Creative Commons Attribution license; the source of each clip is given by its YouTube id. Member clips are redistributed from VideoUFO under its license.