MLSD — Image Copyright Violation Detection

Detect whether a user-uploaded image violates copyright. VLM embedding + ANN lookup is the consensus retrieval approach; the interesting depth lives in adversarial robustness and partial-violation follow-ups. Requirements Functional: given a user-uploaded image, decide whether it infringes any image in a copyrighted-image registry. Pipeline: pretrained VLM (CLIP-style) → image embedding → ANN index over the registry → similarity threshold → ownership check (uploader vs registry owner). Reported follow-ups (must be ready to answer): How would you fine-tune the VLM? (Contrastive learning on copyright-positive pairs, hard-negative mining.) What if the user uploads a 3×3 collage where only one cell infringes? (Image segmentation / region proposals → per-region embedding; or upstream object/region detector before similarity check.) What about adversarial text overlaid on the image? (Contrastive training with adversarial augmentations; OCR-aware preprocessing.) Original is copyrighted, but the user took a phone photo of it — still infringing? (Define legal scope with PM/legal; if yes, training data must include re-photographed pairs as positives.) Operational: registry update pipeline, model-confidence-based human review, per-region thresholds. Notes The candidate who reported this prompt is from search/ads MLE background; the interviewer kept pushing for a single-model solution after the candidate proposed a segmentation + similarity two-model design. Be ready to defend a single-model approach (e.g., a VLM that outputs region-aware embeddings). Adversarial robustness is the second-half emphasis — don't spend all the budget on the base similarity pipeline. Preparation Memorize a CLIP-fine-tuning + ANN retrieval template. Prep a 60-second pitch on contrastive learning with adversarial augmentations. Practice scoping with the interviewer: which infringement types count (exact, transformed, partial, re-captured)?