The invention relates to a global mode and high-frequency residual error collaborative AI generated image model
traceability method. The method mainly comprises the steps that firstly, a global mode
fingerprint branch and high-frequency residual error
fingerprint branch dual-channel collaborative structure is provided according to potential
fingerprint features in an AI generation image, the global mode fingerprint
branch models features such as overall
semantics and style
modes of the image according to
global information in the image, and the high-frequency residual error fingerprint branch is obtained; a lightweight visual Transform structure is used to extract fingerprint features formed by the
generative model globally, and a high-frequency residual fingerprint branch focuses on high-frequency residual features to capture local high-frequency fingerprint features of the
generative model; secondly, a joint training framework based on combination of center loss and
cross entropy loss is used in the training process, and the discrimination capability of the model is enhanced by constraining intra-class feature convergence and inter-class feature separation; and finally, aiming at an
unknown source generation model in practical application, introducing a maximum Softmax probability, an energy
score and the like to realize open set detection, so that the rejection performance of unknown categories is effectively improved. According to the method, complementary modeling is carried out from two angles of global mode fingerprints and local high-frequency residual errors, so that attribution
traceability of models from multi-category AI generated images is realized, generalization and robustness are improved while closed set accuracy is guaranteed, and therefore, the problem of efficient and reliable AI generated image model
traceability in a complex multi-source scene is solved, and the method is suitable for large-scale popularization and application. And the method has a certain application value in actual scenes.