Question 10
IIIAn attacker contributes maliciously crafted data points into a public dataset that a company later uses for training, causing the model to behave incorrectly on certain inputs at inference time. What is this attack called?
Correct answer: B
Explanation
This is data poisoning because the attacker “contributes maliciously crafted data points” into the dataset used for training. Poisoned training data can “corrupt model behavior” and cause the model to act incorrectly on certain inputs at inference time.
Why each option is right or wrong
A. Eavesdropping.
Eavesdropping means secretly intercepting communications or data, not corrupting training examples.
B. Data poisoning, where training data is manipulated to corrupt model behavior.
The described conduct matches an attack on the training set itself: the adversary inserts or alters data before model fitting, so the learned parameters are skewed and the model later misclassifies or behaves erratically on targeted inputs. In ML security taxonomy, this is data poisoning, as opposed to an inference-time attack, because the corruption occurs during training data collection and can persist into deployment.
C. Denial-of-service.
Denial-of-service aims to exhaust availability, not manipulate learned model behavior through training data.
D. Backpropagation interference.
Backpropagation is the learning algorithm step; this is not a standard attack name for poisoned datasets.