ABSTRACT
Machine learning models based on Deep Neural Networks (DNNs) are increasingly deployed in a wide variety of applications, ranging from self-driving cars to COVID-19 diagnosis. To support the computational power necessary to train a DNN, cloud environments with dedicated Graphical Processing Unit (GPU) hardware support have emerged as critical infrastructure. However, there are many integrity challenges associated with outsourcing the computation to use GPU power, due to its inherent lack of safeguards to ensure computational integrity. Various approaches have been developed to address these challenges, building on trusted execution environments (TEE). Yet, no existing approach scales up to support realistic integrity-preserving DNN model training for heavy workloads (e.g., deep architectures and millions of training examples) without sustaining a significant performance hit. To mitigate the running time difference between pure TEE (i.e., full integrity) and pure GPU (i.e., no integrity), we combine random verification of selected computation steps with systematic adjustments of DNN hyperparameters (e.g., a narrow gradient clipping range), which limits the attacker's ability to shift the model parameters arbitrarily. Experimental analysis shows that the new approach can achieve a 2X to 20X performance improvement over a pure TEE-based solution while guaranteeing an extremely high probability of integrity (e.g., 0.999) with respect to state-of-the-art DNN backdoor attacks. © 2022 ACM.