The invention relates to a trusted
deep learning reasoning method for an ARM heterogeneous
edge device. The method comprises the following steps: a deployment stage: storing a structure file and a weight of a model in a core stack of a TEE through a static
trusted application PTA; in the reasoning stage, REE loads a structure model and a weight from TEE secure storage, initializes a reasoning model and creates a
thread pool; dynamic trusted applications TA0 and TA1 are adopted for asynchronous cooperative
verification, TA0 executes non-convolutional layer reasoning and convolutional layer
verification, TA1 executes convolutional layer
verification, and TA0 and TA1 are executed in a staggered mode; verifying a
convolution layer reasoning result by using an FTC
algorithm; the GPU is used for accelerating parallel reasoning of a
convolution layer, multi-TA cooperative verification is executed through the multi-core CPU, and asynchronous parallelism of model credible reasoning and
convolution credible verification is achieved. The method is low in time
delay and high in reliability, the privacy and integrity of the
deep learning model can be protected, the
time overhead of credible reasoning is reduced, and the real-time performance of credible
deep learning reasoning on the ARM heterogeneous
edge device is improved.