System Design: Detect Dangerous Objects in Facebook Ads

How would you design a system that detects dangerous objects in facebook ads?

Asked in the ML system design stage. Discuss an end-to-end system for identifying weapons, explosives, and other prohibited or restricted objects in ad images, video, text, and landing-page media.

Address the following:

1. Clarify policy scope, severity, and latency requirements.
2. Propose ingestion, candidate detection, classification, and human-review stages.
3. Explain online versus batch inference and estimate scale using explicit assumptions.
4. Define labels, hard-negative mining, active learning, and retraining.
5. Describe offline and online evaluation, monitoring, drift detection, and rollback.

Consider multimodal ads, regional policy differences, adversarial edits, new object categories, reviewer capacity, privacy, and the tradeoff between dangerous-content recall and advertiser false positives.