Exploring the landscape of backdoor attacks on deep neural network models

Abstract

Deep neural networks have recently been demonstrated to be vulnerable to backdoor attacks. Specifically, by introducing a small set of training inputs, an adversary is able to plant a backdoor in the trained model that enables them to fully control the model's behavior during inference. In this thesis, the landscape of these attacks is investigated from both the perspective of an adversary seeking an effective attack and a practitioner seeking protection against them. While the backdoor attacks that have been previously demonstrated are very powerful, they crucially rely on allowing the adversary to introduce arbitrary inputs that are -- often blatantly -- mislabelled. As a result, the introduced inputs are likely to raise suspicion whenever even a rudimentary data filtering scheme flags them as outliers. This makes label-consistency -- the condition that inputs are consistent with their labels -- crucial for these attacks to remain undetected. We draw on adversarial perturbations and generative methods to develop a framework for executing efficient, yet label-consistent, backdoor attacks. Furthermore, we propose the use of differential privacy as a defence against backdoor attacks. This prevents the model from relying heavily on features present in few samples. As we do not require formal privacy guarantees, we are able to relax the requirements imposed by differential privacy and instead evaluate our methods on the explicit goal of avoiding the backdoor attack. We propose a method that uses a relaxed differentially private training procedure to achieve empirical protection from backdoor attacks with only a moderate decrease in acccuacy on natural inputs.