Method and system for performance testing of deep learning algorithm in electromagnetic field environment

Abstract

Disclosed are a method and system for performance testing of a deep learning algorithm in an electromagnetic field environment. The method includes: establishing a physical-logical topology mapping relationship between a target hardware physical space and logical computing nodes of a deep learning algorithm; generating a physical characteristic constraint operator based on electromagnetic environment parameters and hardware sensitivity; monitoring an execution state of hardware instructions, and performing error injection on an intermediate-layer tensor by utilizing the physical characteristic constraint operator and the physical-logical topology mapping relationship upon identifying an electromagnetic susceptibility window period of a high-energy-consumption operator; and calculating a performance loss based on an algorithm output, and performing a closed-loop update on the electromagnetic environment parameters using a physical sensitivity coefficient until a critical vulnerable path is locked. The present invention overcomes the related art's inability to simulate hardware-localized physical disturbances that cause structured algorithm failures.