SpecDetect: Simple, Fast, and Training-Free Detection of LLM-Generated Text via Spectral Analysis

TL;DR

SpecDetect reframes LLM-generated text detection as a frequency-domain signal-processing problem by treating the token log-probability sequence as a zero-mean signal and analyzing its spectrum. It identifies DFT Total Energy $E_{DFT}$ as a single, hyperparameter-free, robust discriminant, with SpecDetect++ further boosting robustness via sampling-discrepancy. Empirical results across diverse datasets, models, and decoding strategies show state-of-the-art performance and favorable efficiency, including strong robustness to paraphrasing, varying text lengths, and cross-model/non-English generalization. The work demonstrates that classical signal processing can provide a simple, interpretable, and highly effective pathway for detecting LLM-generated text in practical settings.

Abstract

The proliferation of high-quality text from Large Language Models (LLMs) demands reliable and efficient detection methods. While existing training-free approaches show promise, they often rely on surface-level statistics and overlook fundamental signal properties of the text generation process. In this work, we reframe detection as a signal processing problem, introducing a novel paradigm that analyzes the sequence of token log-probabilities in the frequency domain. By systematically analyzing the signal's spectral properties using the global Discrete Fourier Transform (DFT) and the local Short-Time Fourier Transform (STFT), we find that human-written text consistently exhibits significantly higher spectral energy. This higher energy reflects the larger-amplitude fluctuations inherent in human writing compared to the suppressed dynamics of LLM-generated text. Based on this key insight, we construct SpecDetect, a detector built on a single, robust feature from the global DFT: DFT total energy. We also propose an enhanced version, SpecDetect++, which incorporates a sampling discrepancy mechanism to further boost robustness. Extensive experiments show that our approach outperforms the state-of-the-art model while running in nearly half the time. Our work introduces a new, efficient, and interpretable pathway for LLM-generated text detection, showing that classical signal processing techniques offer a surprisingly powerful solution to this modern challenge.

Figures (11)

• Figure 1: Comparison of a representative human vs. LLM-generated sample (XSum dataset, LLaMA3-8B as source model, GPT-J-6B as proxy model), showing both time-domain signals and the frequency-domain spectrum.
• Figure 2: The overall framework of the SpecDetect method. The process consists of three main stages: (1) Inference to obtain the token probability sequence, (2) Frequency-domain transformation to compute the DFT Total Energy, and (3) Scoring to make a final classification.
• Figure 3: Distributions of spectral features for human and LLM-generated text across different datasets and source models. The top row displays energy-based features ($E_{DFT}$, $E_{STFT}$, $\overline{F}_{spec}$), while the bottom row shows frequency-based features ($C_{spec}$, $H_{spec}$, $V_{H_{spec}}$). The visualization clearly demonstrates that energy-based features provide a strong and consistent separation between the two classes, whereas the distributions for frequency-based features show significant overlap.
• Figure 4: Efficiency comparison of detection methods. Runtimes are averaged per sample (in milliseconds) and labeled on each bar. The bars for perturbation-based methods with extremely long runtimes have been truncated.
• Figure 5: Performance vs. Efficiency trade-off. The x-axis represents detection performance (AUC), and the y-axis represents average runtime (in milliseconds). Methods are grouped by category using marker shapes, and our proposed methods are highlighted in red. Several methods (i.e., Entropy, DetectGPT, DetectNPR, DNA-GPT) are excluded as outliers due to non-competitive performance or prohibitive runtimes. The coordinates for Lastde++ are explicitly labeled because its runtime exceeds the y-axis range.