What is a backdoor attack in ML and how does trigger-based activation differ from natural triggers?

• A. Backdoor Persists Via Hidden Behavior Triggered By A Specific Input; Triggers Can Be Explicit (Pattern) Or Subtle
• B. Backdoors Are The Same As Data Poisoning
• C. Triggers Must Be Explicit Patterns Only
• D. Backdoors Always Degrade Performance

Answer: (A)

A backdoor attack in machine learning is when a model is trained to hide a specific behavior that only activates when a particular trigger appears in the input. The model behaves normally on inputs without the trigger, but when the trigger is present, it outputs a pre-determined, attacker-chosen result. Those triggers can be explicit patterns, like a visible sticker or patch, or they can be subtle perturbations that are hard to notice. The idea is that the backdoor persists inside the model’s parameters, so the malicious behavior remains even after standard evaluation on clean data.

The difference between trigger-based activation and natural triggers is that triggers are deliberately crafted signals meant to reliably flip the model to the attacker’s target, whereas natural triggers would be incidental features or variations that occur in data and aren’t intended to produce a controlled outcome. Importantly, backdoors don’t necessarily degrade overall performance on clean inputs; a backdoored model can still achieve high accuracy while secretly responding to inputs containing the trigger in a way chosen by the attacker.