Katkat
Back

Build and evaluate illegal-video classifier

GoogleData Scientist · Technical Screen
HardData ScientistTechnical ScreenMachine Learning

Last updated

This question tests the ability to architect an end-to-end classification pipeline for high-stakes, multimodal content moderation while accounting for class imbalance, adversarial evasion, and operational safety.

Problem Statement

You are tasked with building a production system to detect a specific category of illegal/prohibited video content at scale (e.g., millions of uploads per day). The platform relies on user-generated content, which is characterized by extreme class imbalance (the illegal content is <0.01% of total traffic), noisy labels, and adversarial actors who use transformations (e.g., blurring, speed shifts, audio overlays) to bypass detection.

Requirements

  1. Multimodal Architecture: Describe how to combine visual frames, audio tracks, and associated metadata/text to classify a video. How would you handle feature fusion?
  2. Training Strategy: Propose a strategy for training on highly imbalanced, sparse, and potentially mislabeled data. How do you handle 'hard negatives' and adversarial samples?
  3. Human-in-the-Loop: Design a triage system. Given limited moderator bandwidth, how do you decide which videos require human review versus automated rejection?
  4. Operational Metrics: Beyond traditional accuracy/AUC, what metrics would you track to monitor model degradation, 'false positive' impact on user experience, and adversarial drift?

Constraints

  • The system must provide a low-latency decision (e.g., under 500ms for initial screening).
  • Privacy constraints prevent storing PII or keeping non-violating content for long durations.

Example Scenario

  • Input: A video segment (audio/video file) + uploader metadata (account age, location, history).
  • Output: A probability score and a recommendation (Approve, Block, or Flag for Human Review).

Hints

Hint 1

Consider using a hierarchical approach where a lightweight 'screening' model filters the majority of benign content before invoking a heavy-duty multimodal transformer.

Hint 2

Think about how active learning or semi-supervised learning could help improve the model using the stream of human moderation decisions.

Hint 3

Focus on robust feature engineering; rather than raw pixels, consider embeddings that are invariant to common transformations like re-encoding or resolution downscaling.

Follow-up questions

  • How would you handle an adversarial attack where the perpetrator slightly modifies videos in a way that is imperceptible to humans but causes your model to misclassify them?
  • If the model shows a sudden spike in false positives for a specific geographic region, what is your diagnostic process to isolate if this is a data bias or a system-level issue?
  • How does your system design change if you are required to comply with 'Right to be Forgotten' regulations, necessitating the deletion of training data points upon user request?