Strongly-transferring memorized examples in deep neural networks

Abstract

Training deep neural networks requires large quantities of labeled training data, on the order of thousands of examples per class. These requirements make model training both time-consuming and expensive, which provides an incentive for adversaries to steal, or copy, other users' models. In this work, we examine a recent defense method called neural network watermarking via memorized examples, where an owner intentionally trains his model to mislabel particular inputs. We try to isolate the mechanism by which memorized examples are learned by a model in order to better evaluate their robustness. We find that memorized examples are indeed strongly embedded in trained models and actually transfer to stolen models under one form of model stealing. When access to local input-logit gradient information is used by an attacker, the stolen model also learns to mislabel the memorized examples. We show that this transfer is robust to architecture mismatch and perturbations of the query set used for stealing. We present different possible mechanisms for memorized example transfer and find that local input geometry is insufficient to explain the phenomenon. Finally, we describe a simple method for a model owner to boost the transfer rate of memorized examples, increasing their effectiveness as a defense against model stealing.