farouk04/vigilvid-research

VigilVid Research Dataset

Dataset Summary

Private research dataset for evaluating VigilVid's video-only AI-generated / deepfake detection workflow.

This dataset is intended for academic evaluation, model improvement, and FYP reporting. It should remain private until consent, licensing, and public release conditions are reviewed.

Dataset Structure

data/
  manifest.jsonl
  predictions/
  evaluations/
videos/
  train/
  validation/
  test/
  demo/

Each manifest.jsonl row follows vigilvid-research-v1.

Important fields:

• sample_id
• split
• label
• video_sha256
• hf_video_path
• source_dataset
• license
• consent_scope
• ai_probability
• prediction_label

Labels

• real: authentic / not AI-generated ground truth
• fake: AI-generated / deepfake ground truth

Evaluation

Report at minimum:

• accuracy
• fake precision
• fake recall
• fake F1
• real precision
• real recall
• balanced accuracy
• confusion matrix

Privacy And Consent

Videos are included only when the source dataset license or user opt-in permits research use.

Detection history is metadata only. Raw-video retention is separately controlled by research consent.

Known Limitations

• Binary real/fake labels only.
• No artifact-category classifier in v1.
• Model probabilities are estimates, not proof.
• Dataset composition may not represent all social platforms or manipulation methods.