It appears that any machine learning model can be backdoored during training. The backdoor makes it possible to fool the model during real use.