Video Frame-Sampling Budgets and What They Trade Away

Your platform processes 5 million video uploads/day, averaging 60 seconds each. Scoring every frame at 30 fps would mean roughly 9 billion frame inferences/day. Your team proposes sampling one frame every 3 seconds (≈20 frames per video), escalating to the expensive temporal/audio-visual model only if a sampled frame or account metadata looks suspicious.

1. Estimate, at a high level, the compute reduction this sampling strategy buys versus scoring every frame, and state your assumptions explicitly.
2. Describe a concrete manipulation that this sampling strategy could miss, and why.
3. Propose one concrete way to reduce that risk without abandoning sampling entirely, and explain the trade-off it makes.