DWT-Fusion: Detecting AI-Generated Text via Wavelet Analysis of Token Probabilities

Forum topic · ✨步子哥 · 2026-08-03

Summary

DWT-Fusion is a training-free framework that detects LLM-generated text by treating token log-probabilities from a proxy model as a one-dimensional signal and analyzing it with the Discrete Wavelet Transform (DWT). Instead of relying on global average probability, it captures local and multi-scale probability fluctuations that distinguish AI text from human text, extracting three wavelet-domain scores (first-level detail energy, multilevel detail energy, and window-energy variability). A calibration-guided voting fusion over multiple wavelet configurations further boosts performance. On HC3 the method reaches AUROC 0.9919, on M4 0.8477, and on MAGE 0.7471, all without labeled training data. A DFT baseline underperforms, confirming the value of joint time-frequency information. Limitations include weaker cross-generator results, proxy-model dependency, untested adversarial robustness, and performance still trailing top supervised detectors.

What is DWT-Fusion?

DWT-Fusion is a training-free framework for detecting AI-generated text. It treats the sequence of token log-probabilities produced by a proxy language model (e.g., GPT-2 Medium) as a one-dimensional signal and applies the Discrete Wavelet Transform (DWT) to extract detection features.

Reference: Özdaş, M. B. & Osmanoğlu, M. *DWT-Fusion: A Signal-Based Framework for Training-Free LLM-Generated Text Detection*. arXiv: https://arxiv.org/abs/2607.22026 (2025).

The core problem it addresses

A naive approach is to compute the average log-probability of a text: AI text tends to have higher (smoother) probabilities, human text lower (rougher) probabilities. However, global averaging hides local differences:

A skilled human writer can produce fluent prose with high probabilities.
An AI deliberately imitating human roughness can suppress its own probability signal.

DWT-Fusion's insight is to look at local and multi-scale probability fluctuations, not global averages.

How it works

1. Signal construction. Compute conditional log-probabilities of each token using a proxy open-source LM. The resulting sequence is treated as a 1-D signal (analogous to an audio waveform). 2. Wavelet decomposition. Apply DWT to split the signal into approximation coefficients (low-frequency, global trend) and detail coefficients (high-frequency, local variation) across multiple scales.

Why wavelets and not, say, global statistics or Fourier analysis?

Global statistics lose local information.
DFT (Discrete Fourier Transform) gives frequency content but no localization.
DWT provides joint time-frequency representation, capturing both *what frequency* and *where* in the token sequence.

AI and human texts differ not just globally but in short-range smoothness and longer-range rhythmic patterns. Wavelets expose both.

Three wavelet-domain scores

The framework defines three scalar features:

1. First-level detail energy — fine-scale local fluctuation strength; AI text tends to be more uniform at this scale. 2. Multilevel detail energy — sum of detail energy across all decomposition levels; captures multi-scale volatility. 3. Window-energy variability — variability of energy across sliding windows; captures local heterogeneity.

Each score can be used alone or combined.

Calibration-guided voting fusion

Multiple wavelet configurations (different wavelet families and decomposition levels) each produce a vote. Calibration weights — derived from validation performance — combine these votes, analogous to ensemble learning where stronger base detectors receive higher weight.

Experimental results

Datasets tested:

HC3 — Chinese, relatively simple.
M4 — multi-generator, multi-domain, medium difficulty.
MAGE — multi-generator, multi-domain, multi-language, hardest.

Best single-score AUROC:

| Dataset | AUROC | |---|---| | HC3 | 0.9872 | | M4 | 0.8185 | | MAGE | 0.7138 |

With calibration-guided voting fusion:

| Dataset | AUROC | |---|---| | HC3 | 0.9919 | | M4 | 0.8477 | | MAGE | 0.7471 |

All results are achieved without any labeled training data — only an open-source proxy model is required.

Comparison with DFT baseline

A DFT-based variant consistently underperforms DWT-Fusion across all datasets. This isolates the contribution: it is not signal processing per se that helps, but the joint time-frequency structure of the wavelet transform.

Engineering significance

1. Zero training cost. No labels, no classifier training, only an open-source proxy model. 2. Interpretability. Each of the three scores has a clear physical meaning (local volatility, multi-scale volatility, window heterogeneity), unlike black-box classifiers. 3. Configurability. Wavelet family and decomposition depth can be tuned per deployment scenario.

Honest limitations

MAGE AUROC of 0.7471 implies a non-trivial false-positive rate; insufficient for high-stakes settings such as academic-integrity enforcement.
Proxy-model dependency. Token distributions from a small model (e.g., GPT-2) may not discriminate text generated by much larger models (e.g., GPT-4). The paper studies proxy-model sensitivity but does not test against GPT-4-generated text.
Unverified adversarial robustness. A generator aware of wavelet-based detection could deliberately craft text with human-like multi-scale volatility patterns. No such adversarial test is reported.
Gap to supervised SOTA. Top supervised detectors (RADAR, Ghostbuster) reach AUROC above 0.85 on M4; DWT-Fusion reaches 0.8477 — close but not surpassing. Whether the training-free advantage compensates depends on the deployment context.

One-line takeaway

DWT-Fusion turns token probabilities into a 1-D signal and uses the Discrete Wavelet Transform to expose local, multi-scale volatility patterns that distinguish AI text from human text — no training, no labels, only an open-source model, approaching supervised-detector performance.

---

Paper: https://arxiv.org/abs/2607.22026
HTML: https://arxiv.org/html/2607.22026v1
Code: not yet released