Reliably Bounding False Positives: A Zero-Shot Machine-Generated Text Detection Framework via Multiscaled Conformal Prediction

TL;DR

This work tackles the societal risk of false positives in machine-generated text detection by introducing Multiscaled Conformal Prediction (MCP), a zero-shot framework that bound $FPR$ via length-aware, multiscaled quantiles while preserving detection performance. MCP integrates a flexible nonconformity score derived from a detector and uses length-based calibration to mitigate the degradation seen with traditional CP. The RealDet dataset provides broad domain, multilingual, and adversarially augmented texts to enable realistic calibration and robust evaluation across 22 LLMs. Empirical results show MCP consistently enforces the $FPR$ bound, improves detection performance at low $FPR$ levels, and enhances robustness to adversarial attacks, outperforming traditional calibration methods. These contributions advance reliable, scalable MGT detection suitable for real-world deployment with strict false-positive constraints.

Abstract

The rapid advancement of large language models has raised significant concerns regarding their potential misuse by malicious actors. As a result, developing effective detectors to mitigate these risks has become a critical priority. However, most existing detection methods focus excessively on detection accuracy, often neglecting the societal risks posed by high false positive rates (FPRs). This paper addresses this issue by leveraging Conformal Prediction (CP), which effectively constrains the upper bound of FPRs. While directly applying CP constrains FPRs, it also leads to a significant reduction in detection performance. To overcome this trade-off, this paper proposes a Zero-Shot Machine-Generated Text Detection Framework via Multiscaled Conformal Prediction (MCP), which both enforces the FPR constraint and improves detection performance. This paper also introduces RealDet, a high-quality dataset that spans a wide range of domains, ensuring realistic calibration and enabling superior detection performance when combined with MCP. Empirical evaluations demonstrate that MCP effectively constrains FPRs, significantly enhances detection performance, and increases robustness against adversarial attacks across multiple detectors and datasets.

Figures (10)

• Figure 1: Detection performance of detectors under different framework configurations.
• Figure 2: The MCP Framework. The prediction process consists of four parts, which are executed sequentially: data preparation, nonconformity score definition, multiscaled quantiles calculation, and MGT Detection.
• Figure 3: Left: True Positive Rate (TPR) of different detectors with the CP as a function of $\alpha$. Right: Quantile values calculated for different text length intervals.
• Figure 4: The FPR of various detectors within the MCP framework across all datasets, after applying alpha constraints with values of alpha set to 0.2, 0.1, 0.05, 0.02, 0.01, and 0.005.
• Figure 5: Local ROC curves (with the horizontal axis representing 1 - FPR) for the basic detectors (Binoculars, Fast-DetectGPT) under different real-world attacks, both with and without the MCP framework.

Theorems & Definitions (3)

• Theorem 1
• Corollary 1
• proof