基于编码器-解码器的水印信息嵌入和提取方案OA

Deep model watermarking technology,as an effective method for intellectual property protection,has attracted increasing attention.How to design an effective and reliable watermarking scheme remains a critical challenge in the field of deep model copyright protection.In this context,a black-box watermark embedding and extraction scheme is explored to provide a reliable solution for safeguarding deep learning models.The main contributions are as follows:a watermark embedding and extraction framework based on an encoder-decoder architecture is constructed.In the embedding stage,an embedding-discriminator network is designed,which consists of an embedding network and a discriminator.The initial trigger set is processed by the embedding network first,and the output is evaluated by the discriminator.The discriminator's feedback is then used to refine the embedding network,thereby improving the quality of the embedded watermark.After the embedded trigger set is input into the target model,if the output matches the target category,the watermark can be further extracted.Correspondingly,a symmetric extraction network is designed to mirror the embedding process.This extraction network includes an extractor and a discriminator.The extracted watermark information is evaluated by the discriminator and fed back to the extractor to further enhance the extraction performance.Experiments are conducted to investigate the impact of different frequency domains and training epochs on the extraction performance.Experimental results on five classic deep learning models demonstrate that the proposed watermarking scheme achieves a structural similarity index of over 0.95 between the embedded and extracted watermark,with a watermark extraction success rate consistently exceeding 98％,indicating the effectiveness of the proposed method for watermark verification.