Resilient and secure deep learning-oriented microarchitectures

Abstract

Deep neural networks (DNNs) have emerged as crucial methods for solving complex problems in various domains, including computer vision, natural language processing, and recommendation systems. An ideal DNN-based system should accurately make predictions or classifications based on some input data with no interference from the external environment. However, DNN-based systems are susceptible to failures due to reliability and security issues. This thesis evaluates many compressed DNN models under faulty conditions like bit-flips due to transient or permanent faults. Then, an AN-based detection scheme targeting DNN accelerators deployed into safety-critical systems is proposed. Due to the high compliance standards, microarchitectures employed in this type of system must provide a detection capability of 99% of faults. Also, the AN-based detection offers a lightweight solution, particularly when incorporated with the novel AN code-aware quantization technique proposed in this thesis. Training-based obfuscation techniques have been successfully employed to protect DNN models from illegal access to sensitive data, such as the parameters. However, crucial information such as the output class distribution can be leaked to attackers, indicating that the target model has been compromised. Additionally, an obfuscation scheme must provide a scalable way to protect the model providers’ portfolios. The novel swap-based obfuscation scheme provides a robust obfuscation of the DNN model parameters through a scalable and secure solution, avoiding any illegal access and use of the model by non-authorized entities