Browse Teach Curate

Game-Theoretic Model of Black-Box Deep Neural Network Watermarking

A formulation of trigger-based black-box model watermarking as a two-player game between a defender who embeds and verifies an ownership watermark and an attacker who tries to remove or forge it. Each player's payoff function is constructed and their optimal responses derived, giving model watermarking a theoretical foundation beyond trial-based or purely data-driven schemes and guiding the design of new watermarking methods.

B9 Text

A Game Between the Defender and the Attacker for Trigger-based Black-box Model Watermarking Chaoyue

A game-theoretic model of trigger-based black-box watermarking of deep neural networks, where a watermark is embedded by backdooring the model so that specific trigger inputs yield a secret owner-cho…