Secured convolutional layer hardware co-processor in convolutional neural network (CNN) using facial biometric

Abstract

The chapter describes a methodology for designing a secured custom reusable hardware co-processor intellectual property (IP) core for convolutional neural network (CNN)-based convolutional layer [1]. In this methodology, first the convolutional layer hardware IP core is designed using high-level synthesis (HLS) process. HLS transforms the behavioral description/transfer function corresponding to the convolution operation of CNN into scheduled data flow graph (DFG) design to realize the behavior of convolutional layer. Since the reusable IP cores may be vulnerable to the hardware threat of IP piracy, therefore, facial biometric of IP vendor is integrated within the design. To do so, first, the facial signature corresponding to facial biometric image of an IP vendor is generated. Subsequently, facial biometric signature in the form of encoded hardware security constraints (digital evidence) is integrated into the IP design during the register allocation module of HLS. HLS-based design methodology results in a custom reusable convolutional layer hardware co-processor IP core with lesser implementation complexity and robust security. The embedded facial biometric-driven digital evidence enables the robust detective control against IP piracy. Therefore, it ensures the integration of genuine reusable hardware co-processor IP cores in the system-on-chip (SoCs) of consumer electronics (CE) systems, thereby also ensuring the integrity and safety of end consumers.The organization of the chapter is as follows: Section 4.1 summarizes the introduction of the chapter; Section 4.2 discusses the motivation for designing CNN-based custom convolutional layer reusable IP core; Section 4.3 presents the benefits of the methodology to the end consumer; Section 4.4 presents the discussion on similar existing works; Section 4.5 provides background on CNN framework; Section 4.6 presents an overview of the approach; Section 4.7 provides the details of the approach; Section 4.8 provides analysis and discussion; and Section 4.9 concludes the chapter.