Insert Backdoor Trigger
An attacker hides a secret 'key' or signal inside data—like a tiny, invisible pattern in a photo—that tricks an AI into behaving exactly how the attacker wants, often used after they have already secretly tampered with the AI's training.
The process of injecting a specific, often imperceptible, feature into inference-time input data to activate a pre-existing backdoor in a model. This technique relies on the model having been previously compromised via model poisoning to associate the trigger with a malicious output.
The adversarial insertion of a latent, perceptually non-obvious trigger into input data during the inference phase to elicit a targeted misclassification or malicious behavior. This technique functions as the activation mechanism for a backdoor established through prior model poisoning (AML.T0018.000), exploiting the learned association between the trigger feature and the adversary's desired output state.