Which attack type is MOST likely to cause model drift?

Question

Options

• AModel stealing
• BPerfect knowledge
• CData poisoning
• DMembership inference

• A
  8% (3)
• B
  3% (1)
• C
  87% (33)
• D
  3% (1)

Explanation

Data poisoning involves injecting malicious, manipulated, or mislabeled data into the training pipeline. Over time, the model learns from this corrupted data and drifts away from its intended behavior - its predictions become systematically skewed or degraded. Model stealing (A) creates a copy of the model but does not alter it. Perfect knowledge (B) is a white-box attack strategy that exploits model internals but does not modify training. Membership inference (D) determines whether specific records were used in training, a passive information-gathering attack that does not corrupt the model. Only data poisoning directly corrupts the learning process, causing drift.

No community discussion yet for this question.