TempQ-Jail: Query-Constrained Candidate Ranking for Text-to-Video Jailbreak Attacks
TempQ-Jail ranks text-to-video jailbreak candidates under limited queries, raising success on CogVideoX-5B versus prior methods.
TempQ-Jail treats text-to-video jailbreaks as a query-constrained ranking problem, because generation and security checks are too costly to test many candidates. It combines heterogeneous attack mechanisms and estimates each candidate from security-gate passage, dangerous visual generation, preservation of the original intent, and temporal validity. Evaluated on CogVideoX-5B with 70 intents from T2VSafetyBench against six prior methods, it reached TP-ASR of 48.9% at five queries and 65.4% at ten. Those results exceed the strongest baselines by 4.6 and 4.0 percentage points, with AUC-TP of 0.469 and an average of 6.3 queries.
- TempQ-Jail ranks text-to-video jailbreak candidates under a limited query budget.
- Scoring covers gate passage, dangerous visuals, intent preservation, and temporal validity.
- On CogVideoX-5B, success was 48.9% at five queries and 65.4% at ten.
- It beat the strongest baselines by 4.6 and 4.0 points, with average queries of 6.3.
From the original article
1 min read at arxiv.orgExisting text-to-video (T2V) jailbreak methods mainly seek more effective or stealthier attack candidates. In guarded T2V systems, however, video generation and security evaluation are costly, so an attacker often cannot test a large candidate pool. We therefore formulate T2V jailbreak as a query-constrained candidate allocation and ranking problem and propose TempQ-Jail.
By Tianmeng Fang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.