Protecting Deep Cerebrospinal Fluid Cell Image Processing Models with Backdoor and Semi-Distillation

Cerebrospinal fluid image is an informative source for the diagnosis of many diseases. Consequently, deep learning models for cerebrospinal fluid image processing turn out to be a promising computer-aided diagnosis technique. Current models can efficiently and correctly identify numerous categories of cells within an image of cerebrospinal fluid. Training a cerebrospinal fluid image processing model, especially a deep neural network, requires a vast amount of data and computation. Collecting necessary data for medical tasks is an expensive procedure, during which many experts, devices, and privacy concerns are involved. Therefore, it is crucial to protect these deep models from piracy and reselling. In this paper, we study the problem of intellectual property protection of deep cerebrospinal fluid image processing models. We adopt the backdoor-based watermark as the ownership evidence and propose a semi-distillation framework to embed the watermark into the model. The proposed scheme can verify the ownership of the genuine author, hence provide robust and unforgeable protection over deep cerebrospinal fluid image processing models.