Crowdsourced privacy-preserved feature tagging of short home videos for machine learning ASD detection

Editorial reading guide · Respect

What privacy protection would be adequate before sharing my child’s video for AI analysis?

Face masking and pitch shifting preserved high sensitivity in this small sample, with lower specificity under combined alterations. The work makes the privacy–utility trade-off visible without proving anonymity or a clinic-ready autism screen.

Study context

A certified pool of 102 nonexpert crowd workers rated a balanced set of twenty-five autistic and twenty-five neurotypical children’s videos for two previously trained classifiers.

Age / stage: Children under eight years in fifty selected home videos

Read with three questions

1. Which video alterations were tested and what happened to specificity?
2. How were parent-reported diagnoses used as a reference?
3. What privacy risks remain beyond maintaining classifier accuracy?

Keep the limits in view

The selected fifty-video sample, parent-reported diagnoses, binary autism classification, and uncontrolled video differences limit generalization. Accuracy after alteration does not establish comprehensive de-identification, unbiased performance, or readiness for clinical use.