Exploration of Uninitialized Configuration Memory Space for Intrinsic Identification of Xilinx Virtex-5 FPGA Devices

SRAM-based fingerprinting uses deviations in power-up behaviour caused by the CMOS fabrication process to identify distinct devices. This method is a promising technique for unique identification of physical devices. In the case of SRAM-based hardware reconfigurable devices such as FPGAs, the integrated SRAM cells are often initialized automatically at power-up, sweeping potential identification data. We demonstrate an approach to utilize unused parts of configuration memory space for device identification. Based on a total of over 200,000 measurements on nine Xilinx Virtex-5 FPGAs, we show that the retrieved values have promising properties with respect to consistency on one device, variety between different devices, and stability considering temperature variation and aging. 1. Introduction Identification of devices is a primitive that plays a crucial role for a number of applications, including authentication of devices and protection against cloning of devices (cocalled product piracy) or intellectual property. IDs which are stored in nonvolatile memory can often be easily cloned or modified. Hence approaches have been published to overcome the aforementioned drawbacks. They are usually based on unique physical properties of the single chip. For example, such properties are caused by manufacturing process variations. The two main approaches in this context are physical fingerprinting and the use of physical uncloneable functions (PUFs). Former strives to identify a given circuit directly by physical characteristics latter use physical characteristics to perform a challenge-response authentication. A promising technique used for both approaches is to observe the state of uninitialized SRAM cells. When voltage above a certain threshold is applied to an SRAM cell its initial unstable state will change to one of two possible stable states “0” or “1”. The probability for each stable state is heavily dependant on small variations originated during the CMOS fabrication process causing slight deviation in threshold voltage inside the cells. The probability varies between different cells even inside a single chip thus representing a characteristic initial memory content on power-up for each device. Depending on the probability distribution the major part of the memory content is stable for most of the power-ups. Other bits having a probability around 50% show a power-up behaviour similar to random noise. Assuming a high rate of stable data the memory content can be used to provide high quality identification data that is very hard to reproduce deliberately.