Adversarial Attacks on Text-to-Video Models
Analysis
This paper addresses a critical, yet under-explored, area of research: the adversarial robustness of Text-to-Video (T2V) diffusion models. It introduces a novel framework, T2VAttack, to evaluate and expose vulnerabilities in these models. The focus on both semantic and temporal aspects, along with the proposed attack methods (T2VAttack-S and T2VAttack-I), provides a comprehensive approach to understanding and mitigating these vulnerabilities. The evaluation on multiple state-of-the-art models is crucial for demonstrating the practical implications of the findings.
Key Takeaways
- •Introduces T2VAttack, a framework for adversarial attacks on Text-to-Video models.
- •Focuses on both semantic and temporal aspects of video generation.
- •Proposes two attack methods: T2VAttack-S (synonym substitution) and T2VAttack-I (word insertion).
- •Evaluates the adversarial robustness of several state-of-the-art T2V models.
- •Demonstrates that even small prompt modifications can significantly degrade video quality.
“Even minor prompt modifications, such as the substitution or insertion of a single word, can cause substantial degradation in semantic fidelity and temporal dynamics, highlighting critical vulnerabilities in current T2V diffusion models.”