Design an ML system that detects whether a user-uploaded image violates copyright.
Requirements
-
Input: an image uploaded by a user (optionally with text/caption and user/account metadata).
-
Output: a decision such as
{allow, block, human_review}
plus an explanation signal (e.g., matched copyrighted work ID, similarity score, region of match).
-
Must scale to large numbers of uploads and a large database of copyrighted/reference images.
Follow-up questions to address
-
How would you fine-tune
a pretrained vision(-language) model used for embeddings or classification?
-
If an uploaded image is a
3×3 grid (collage)
and only one tile is infringing, how do you detect that?
-
How do you handle
adversarial text overlays
or other adversarial manipulations intended to evade detection?
-
If the original work is copyrighted, does a
photo of the work
(e.g., taken by a camera from a screen/poster) count as infringement, and how should the system treat such near-duplicates?
Assume you can combine retrieval, classification, and human review, and you must justify metrics, thresholds, and latency/cost tradeoffs.