End-side deep learning model protection method and system based on code obfuscation

• Main Authors: YANG XIAOLIN, WANG YANG, SONG FEIYANG, YUAN ZIXIN, HU WEN'AO, YAN FEI, ZHAO XINMIAO
• Format: Patent
• Language: Chinese; English
• Published: 16.07.2024
• Subjects: CALCULATING; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention discloses an end-side deep learning model protection method and system based on code obfuscation, belongs to the technical field of deep learning model protection, and is used for solving the problems that an existing model protection technology aims at protecting intellectual property of a model, cannot substantially prevent the model from being stolen and reconstructed by an attacker, and cannot protect the model. And the requirements of defending model decompilation attacks cannot be met. The method comprises the steps of transmitting a to-be-deployed deep learning model into a deep learning model compiler for optimization processing; performing confusion processing on the optimized to-be-deployed deep learning model through a model protection framework in a deep learning model compiler to obtain an intermediate representation file; wherein the model protection framework comprises an operator confusion module and/or a parameter confusion module and/or a topology confusion module; and compilin