Define a backdoor attack in machine learning and provide a scenario.

• A. It is when a program uses a backdoor to log user data.
• B. A backdoor attack plants a trigger in training so the model behaves maliciously when the trigger appears (e.g., misclassifying images with a specific pattern).
• C. It is when an attacker distorts the output after deployment to degrade performance during peak load.
• D. It is when a model learns to ignore certain classes during training.

Answer: (D)

A backdoor attack in machine learning is when an attacker secretly embeds a trigger into the model during training so that the model behaves in a malicious or attacker-controlled way whenever that trigger appears, while acting normally on other inputs.

For example, an attacker poisons a small fraction of the training data by adding a subtle pattern to images and labeling them as a specific target class. After training, the model will correctly classify normal inputs, but when the same trigger pattern shows up in a new input, the model intentionally outputs the attacker-chosen class. A concrete scenario is a traffic sign recognition system that normally identifies signs correctly but misclassifies a stop sign as a speed-limit sign whenever a small sticker matching the trigger is present.

This explanation aligns with the idea of a hidden, trigger-based behavior that only activates under specific conditions. The other options describe different issues (logging data, degrading performance after deployment, or ignoring a class) that do not involve a covert trigger that alters behavior on demand.