Tags:Backdoor watermark, Deep neural network, Federated learning, Model watermarking and Ownership Verification
Abstract:
Big data has significantly propelled the advancement of artificial intelligence (AI), notably in deep learning domains. Yet, the resource-intensive nature of training deep neural networks (DNNs) underscores the critical need for model protection and ownership assertion. Although neural network model watermarking offers a solution, its applicability is limited in federated learning scenarios. This paper introduces VeriChroma, a pioneering framework designed to safeguard DNN models and establish ownership in these contexts. VeriChroma allows individual clients to independently embed and verify private ID-based watermarks, facilitating straightforward ownership claims. It innovatively addresses client constraint conflicts through image blocking and position mapping, guaranteeing unique watermark integration for each participant. Additionally, VeriChroma employs RGB filters to create watermark triggers, enhancing both robustness and secrecy. Our experimental results validate VeriChroma's efficacy and practicality, demonstrating its superior capability in securing DNN model ownership, mitigating federated learning disputes, and providing robust, discreet watermarking. Ultimately, VeriChroma represents a significant stride toward advanced security and intellectual property protection in federated learning environments.
VeriChroma: Ownership Verification for Federated Models via RGB Filters