Design a copyright detection system for video uploads

What's really being asked

This question probes your ability to design a high-throughput, low-latency content moderation pipeline that is robust to real-world distortions like transcoding, cropping, and compression. The interviewer cares about your familiarity with signal-processing techniques, distributed systems patterns, and policy enforcement at scale.

The full answer

First, an event-driven ingestion pipeline where uploads land in object storage and trigger asynchronous preprocessing workers, keeping the upload path non-blocking. Second, media decomposition into parallel audio and video streams. For video, perceptual hashing algorithms like pHash or block mean value hashing that survive re-encoding and minor edits; for audio, acoustic fingerprinting such as Chromaprint or Echoprint that identifies clips despite bitrate changes or noise. Third, a reference index architecture, typically a sharded database or vector store holding fingerprints of registered works, queried with approximate nearest neighbor search to achieve sub-linear lookup times. Fourth, a policy and enforcement tier that routes matches through confidence scoring, automated actions like block or monetize, and human review queues with appeal workflows.

The mistakes people make

Proposing MD5 or SHA checksums, which fail the moment a file is re-encoded. Suggesting pixel-by-pixel or raw audio waveform comparison, which is brittle and computationally infeasible at scale. Designing a single synchronous monolith that processes the entire file before acknowledging upload, creating a terrible user experience and a horizontal scaling bottleneck. Ignoring the distinction between audio and video fingerprints, or omitting the policy layer entirely.

What usually comes next

How would you handle false positives from short royalty-free samples or fair-use clips? What happens when a content owner disputes a match or a user appeals a block? How do you keep reference index latency low as it grows to billions of fingerprints? How would you architect around regional copyright law differences?

A concrete example

YouTube's Content ID system, implemented around 2007, compares uploaded videos against audio and video files registered by content owners. By 2016 Google had spent roughly 60 million dollars developing it, and the system had facilitated approximately 2 billion dollars in payments to copyright holders. By 2018 investment exceeded 100 million dollars, underscoring that production-grade fingerprinting requires massive ongoing infrastructure in ingestion, indexing, and policy enforcement.