Model poisoning
Also known as: training data poisoning · data poisoning
Model poisoning is tampering with the training or fine-tuning data so the model learns a hidden, malicious behavior — a backdoor.
- A few poisoned examples can trigger a failure on a specific input.
- Hard to detect: the model looks normal otherwise.
- A real concern for any pipeline that ingests external data.
The threat model: if anyone can inject data into your training/fine-tune set, they can potentially plant a backdoor. Control who and what enters the data pipeline.