pretrained model manipulation

We introduce a method to manipulate neuron activations while pre-training models, allowing highly successful inference of sensitive properties of the victim’s downstream training data.

We introduce a method to manipulate neuron activations while pre-training models, allowing highly successful inference of sensitive properties of the victim’s downstream training data.

We present a novel threat model in which adversaries inject stealthy backdoors that activate only when the model is compressed.

We present a novel threat model in which adversaries inject stealthy backdoors that activate only when the model is compressed.