ABSTRACT
This chapter demonstrates how an attacker can learn and replicate the proprietary algorithm. It provides a secure memristor-based neuromorphic computing system (MNCS) design to thwart such replication attacks by leveraging memristors obsolescence effect. In 2008, HP Lab first reported their discovery of a nanoscale memristor based on TiO2 thin-film devices. An MNCS consists of the following proprietary information: Training data denotes the sample set used for training the MNCS and Learning model denotes the model that has been trained for the proprietary application using the training data. In order to solve the above dilemma, in this proposal, we propose to design an MNCS that has a very nonlinear degradation in accuracy. Increasing the depth of neural network is another way to increase the nonlinearity of the MNCS. Each operation cycle of the proposed secured MNCS consists of two periods that is, evaluating operations and system calibration.
