A Self-Oscillated Organic Synapse for In-Memory Two-Factor Authentication

Entering the era of AI 2.0, bio-inspired target recognition facilitates life. However, target recognition may suffer from some risks when the target is hijacked. Therefore, it is significantly important to provide an encryption process prior to neuromorphic computing. In this work, enlightened from time-varied synaptic rule, an in-memory asymmetric encryption as pre-authentication is utilized with subsequent convolutional neural network (ConvNet) for target recognition, achieving in-memory two-factor authentication (IM-2FA). The unipolar self-oscillated synaptic behavior is adopted to function as in-memory asymmetric encryption, which can greatly decrease the complexity of the peripheral circuit compared to bipolar stimulation. Results show that without passing the encryption process with suitable weights at the correct time, the ConvNet for target recognition will not work properly with an extremely low accuracy lower than 0.86%, thus effectively blocking out the potential risks of involuntary access. When a set of correct weights is evolved at a suitable time, a recognition rate as high as 99.82% can be implemented for target recognition, which verifies the effectiveness of the IM-2FA strategy. In-memory two-factor authentication (IM-2FA) is implemented to enhance the security of artificial neural networks. Upon successful completion of the initial authentication, where a set of accurate weights is verified, subsequent target recognition is authorized with an impressive recognition rate of 99.82%. However, in cases where incorrect weights are provided for neuromorphic computing, access is denied. image