GuardianMPC: Backdoor-Resilient Neural Network Computation

Abstract

The rapid growth of deep learning (DL) has raised serious concerns about users' data and neural network (NN) models' security and privacy, particularly the risk of backdoor insertion when outsourcing the training or employing pre-trained models. To ensure resilience against such backdoor attacks, this work presents GuardianMPC, a novel framework leveraging secure multiparty computation (MPC). GuardianMPC is built upon garbled circuits (GC) within the LEGO protocol framework to accelerate oblivious inference on FPGAs in the presence of malicious adversaries that can manipulate the model weights and/or insert a backdoor in the architecture of a pre-trained model. In this regard, GuardianMPC is the first to offer private function evaluation in the LEGO family. GuardianMPC also supports private training to effectively counter backdoor attacks targeting NN model architectures and parameters. With optimized pre-processing, GuardianMPC significantly accelerates the online phase, achieving up to 13.44× faster computation than its software counterparts. Our experimental results for multilayer perceptrons (MLPs) and convolutional neural networks (CNNs) assess GuardianMPC's time complexity and scalability across diverse NN model architectures. Interestingly, GuardianMPC does not adversely affect the training accuracy, as opposed to many existing private training frameworks. These results confirm GuardianMPC as a high-performance, model-agnostic solution for secure NN computation with robust security and privacy guarantees.