Data poisoning, model poisoning and backdoors
A · 10-second intuition Poisoning attacks change what a model learns (or what it retrieves) so that it later misbehaves, sometimes only when a secret trigger appears.
Predict before you read
A model behaves normally but outputs attacker text whenever a rare phrase appears. What is this?