Does DetectGPT Fully Utilize Perturbation? Bridging Selective Perturbation to Fine-tuned Contrastive Learning Detector would be Better
This work addresses the need for more robust and generalizable detectors for machine-generated text, which is crucial for mitigating abuse of large language models, but it is incremental as it builds upon existing perturbation-based methods.
The paper tackles the problem of detecting machine-generated text by addressing limitations in DetectGPT's random perturbation strategy and threshold dependency, proposing a fine-tuned detector called Pecola that uses selective perturbation and contrastive learning, resulting in an average accuracy improvement of 1.20% over state-of-the-art methods on four public datasets.
The burgeoning generative capabilities of large language models (LLMs) have raised growing concerns about abuse, demanding automatic machine-generated text detectors. DetectGPT, a zero-shot metric-based detector, first introduces perturbation and shows great performance improvement. However, in DetectGPT, the random perturbation strategy could introduce noise, and logit regression depends on the threshold, harming the generalizability and applicability of individual or small-batch inputs. Hence, we propose a novel fine-tuned detector, Pecola, bridging metric-based and fine-tuned methods by contrastive learning on selective perturbation. Selective strategy retains important tokens during perturbation and weights for multi-pair contrastive learning. The experiments show that Pecola outperforms the state-of-the-art (SOTA) by 1.20% in accuracy on average on four public datasets. And we further analyze the effectiveness, robustness, and generalization of the method.