End-to-end Customization of Efficient, Private, and Robust Neural Networks

Advancements in machine learning (ML) algorithms, data acquisition platforms, and high-end computer architectures have fueled an unprecedented industrial automation. An ML algorithm captures the dynamics of a task by learning an abstract model from domain-specific data. Once the model is trained by the ML algorithm, it can perform the underlying task with relatively high accuracy. This thesis is specifically focused on Deep Neural Networks (DNNs), a modern class of ML models that have shown promising performance in various applications. Thanks to DNNs, the breadth of automation has been expanded to tasks that were formerly too complex to be performed by computers; nowadays DNNs establish the foundation of applications such as voice recognition, medical image analysis, face authentication, to name a few.Despite DNNs' benefits, their deployment in real-world applications may be circumscribed by several factors. First, DNNs are computationally complex and their efficient execution on resource-constrained edge devices is a critical challenge. Second, users of DNN-based applications are often required to expose their data to the service provider, which may violate their privacy. Third, DNN models may fail to function correctly in the presence of malicious attackers. Having the aforementioned challenges in mind, it is a paramount challenge to design DNN-based systems that are efficient to execute, ensure users' privacy, and are robust to malicious attacks.This dissertation provides holistic customization techniques that pave the way for efficient, private, and robust DNN inference. The key contributions of the thesis are as follows:Efficiency: Development of encoded DNNs, a new family of memory-efficient neural networks. The thesis author's contributions provide customization techniques that enable incorporation of nonlinear encoding to the computation flow of neural networks. An end-to-end framework is introduced to facilitate encoding, bitwidth customization, fine-tuning, and implementation of neural networks on FPGA platforms. Efficiency: Introducing the concept of lookup-table based execution of encoded neural networks. The proposed method replaces floating-point multiplications with look-up table search. A memory-based hardware architecture is then proposed to execute the lookup-based multiplications and accelerate encoded DNN inference. Privacy: Establishing customized solutions for oblivious inference, where a client holds a data sample and a server holds a DNN model. After running the oblivious inference protocol, the client receives the inference result without revealing her input to the server. This thesis proposes automated customization solutions to speed up the oblivious inference while maintaining a high inference accuracy. Robustness: Development of solutions for online detection of neural Trojan triggers, a class of malicious attacks that cause a DNN to perform faulty inferences. The thesis proposes a novel methodology that enhances robustness to Trojan attacks by leveraging dictionary learning and sparse approximation