EURASIP Journal on Information Security, Journal Year: 2024, Volume and Issue: 2024(1)
Published: Oct. 15, 2024
Deep neural networks (DNNs) are fundamental to modern applications like face recognition and autonomous driving. However, their security is a significant concern due various integrity risks, such as backdoor attacks. In these attacks, compromised training data introduce malicious behaviors into the DNN, which can be exploited during inference or deployment. This paper presents novel game-theoretic approach model interactions between an attacker defender in context of DNN attack. The contribution this multifaceted. First, it models interaction using framework. Second, designs utility function that captures objectives both parties, integrating clean accuracy attack success rate. Third, reduces game two-player zero-sum game, allowing for identification Nash equilibrium points through linear programming thorough analysis strategies. Additionally, framework provides varying levels flexibility regarding control afforded each player, thereby representing range real-world scenarios. Through extensive numerical simulations, demonstrates validity proposed identifies insightful guide players following optimal strategies under different assumptions. results indicate fully defense capabilities not always strategy either party. Instead, attackers must balance inducing errors minimizing information conveyed defender, while defenders should focus on risks preserving benign sample performance. These findings underscore effectiveness versatility approach, showcasing across scenarios highlighting its potential enhance against
Language: Английский