Neural Networks for Authenticating Integrated Circuits Based on Intrinsic Nonlinearity

This work presents a machine learning approach to identify integrated circuits based on intrinsic nonlinearity arising out of random variations introduced during device fabrication. The random variations ensure that each integrated circuit has a distinct nonlinearity signature which can be analyzed by a machine learning model to distinguish between chips fabricated from the same mask. We have analyzed multiple samples of two analog-to-digital converters (ADCs) - a continuous-time ΔΣ oversampled ADC and a discrete-time nyquist ADC. The two ADCs have different dominant nonlinearity contributors - inter-symbol interference for the oversampled ADC and static mismatch for the nyquist ADC. A 3-layer artificial neural network can identify the different sample chips for each ADC with a worst-case mean accuracy of 95.97%.