Are you copying my prompt? Protecting the copyright of vision prompt for VPaaS via watermarking

Abstract

Visual Prompt Learning (VPL) reduces resource consumption by avoiding updates to pre-trained model parameters and instead learns input perturbations, a visual prompts, added to downstream task data for predictions. Designing high-quality prompts requires significant expertise and time-consuming optimization, leading to the emergence of Visual Prompts as a Service (VPaaS), where developers monetize well-crafted prompts by providing them to authorized customers.However, in cloud computing environments, prompts can be easily copied and redistributed, posing serious risks to the intellectual property (IP) of VPaaS developers.

To address this, we propose WVPrompt, the first method for protecting visual prompts via watermarking in a black-box setting. WVPrompt consists of two components: prompt watermarking and prompt verification. Specifically, it utilizes a poison-only backdoor attack method to embed a watermark into the prompt, and then employs a hypothesis-testing approach for remote verification of prompt ownership. Extensive experiments were conducted on three well-known benchmark datasets and three popular pre-trained models: RN50, BiT-M, and Instagram. The experimental results demonstrate that WVPrompt is efficient, harmless, and robust to various adversarial operations, making it a reliable solution for securing visual prompts in cloud-based applications.