TRIDENT
An AI-generated text detection framework based on round-trip translation (RTT) that identifies AI text by comparing the semantic consistency between the original text and its back-translation.
Timeline (last 90 days)
TRIDENT is an AI text detection framework based on round-trip translation (RTT), whose core assumption is that AI-generated text and human text differ in semantic consistency.
Experimental results show that all three TRIDENT variants outperform traditional baselines and supervised learning baselines.
TRIDENT was evaluated on seven languages, and adversarial settings and unseen LLMs were also tested.
TRIDENT is sensitive to the translation model, and translation quality directly affects detection performance.
If attackers understand the detection mechanism, they may evade detection by manipulating the translation stability of the text.
Detection methods based on intrinsic language features may have greater long-term value than fingerprint-style detection targeting specific models.
All Facts (6)
TRIDENT is an AI text detection framework based on round-trip translation (RTT), whose core assumption is that AI-generated text and human text differ in semantic consistency.
50%UnverifiedExperimental results show that all three TRIDENT variants outperform traditional baselines and supervised learning baselines.
50%UnverifiedTRIDENT was evaluated on seven languages, and adversarial settings and unseen LLMs were also tested.
50%UnverifiedTRIDENT is sensitive to the translation model, and translation quality directly affects detection performance.
50%UnverifiedDetection methods based on intrinsic language features may have greater long-term value than fingerprint-style detection targeting specific models.
50%UnverifiedIf attackers understand the detection mechanism, they may evade detection by manipulating the translation stability of the text.
50%