DWT-Fusion: Training-Free AI-Generated Text Detection via Wavelet Transforms of Token Probabilities

Forum topic · ✨步子哥 · 2026-07-27

Summary

DWT-Fusion, a paper by Mehmet Batuhan Özdaş and Murat Osmanoğlu of Ankara University (arXiv:2607.22026), treats token-level log probabilities from a proxy language model as a one-dimensional signal and applies the Discrete Wavelet Transform (DWT) to detect LLM-generated text without any training or labeled data. The method derives three interpretable scores—first-level detail energy, multilevel detail energy, and window-energy variability—capturing local and multi-scale probability fluctuations that global averaging misses. A calibration-guided voting fusion combines scores across multiple wavelet configurations. Evaluated training-free on HC3, M4, and MAGE, it reaches AUROC of 0.9919, 0.8477, and 0.7471 respectively, outperforming a DFT-based baseline and approaching supervised detectors like RADAR and Ghostbuster. The post also discusses limitations, including weak performance on multilingual MAGE, proxy-model dependency, and unverified robustness to adversarial attacks.

DWT-Fusion: Catching AI-Generated Text with Wavelet Transforms — Treating Token Probabilities as Signals

> Paper: DWT-Fusion: A Signal-Based Framework for Training-Free LLM-Generated Text Detection > Authors: Mehmet Batuhan Özdaş, Murat Osmanoğlu > Institution: Ankara University, Cyber Security Vocational School > arXiv: 2607.22026 (July 24, 2025)

A Scenario

You receive a student essay and want to know whether it was written by the student or generated by ChatGPT.

You have an open-source language model (e.g., GPT-2 or LLaMA) and can compute the log probability of every token in the text. But how do you use these probabilities to judge authorship?

The most intuitive approach: compute the average log probability over the whole text. AI-generated text usually has higher probability (more "fluent"), while human text usually has lower probability (less "fluent").

But this has a fatal flaw: a skilled human writer's fluent prose also scores high; AI text deliberately mimicking human "disfluency" may score low. Global averaging hides local differences.

DWT-Fusion's core insight: don't look at the global average — look at local, multi-scale probability fluctuations.

Turning Text into a Signal

The first step converts text into a one-dimensional signal.

Concretely: a proxy language model (e.g., GPT-2 Medium) computes the conditional log probability of each token. This sequence of numbers is a one-dimensional signal — like an audio waveform.

Then a Discrete Wavelet Transform (DWT) is applied to this signal.

Why Wavelet Transforms?

The wavelet transform is a classic signal-processing tool. Its core idea: decompose a signal into "details" at different scales.

Analogy: looking at a photo. Global statistics (mean brightness, variance) tell you overall lightness but no details. The Fourier transform tells you which frequencies exist, but not where they occur. The wavelet transform tells you both "which frequency" and "where" — it carries time-domain (position) and frequency-domain (scale) information simultaneously.

DWT-Fusion uses wavelets because AI-vs-human differences may not be global but local and multi-scale:

Local differences: AI may be especially "fluent" (high, stable probability) in some passages while deliberately "disfluent" elsewhere to mimic humans. Global averaging cancels these differences out.
Multi-scale differences: AI probability fluctuations may be small at short scales (a few tokens) but show regular patterns at long scales (dozens of tokens). Human texts fluctuate differently.

The wavelet transform decomposes the signal into "approximation coefficients" (low-frequency, global trend) and "detail coefficients" (high-frequency, local variation) across levels — exactly capturing both kinds of differences.

Three Wavelet-Domain Scores

DWT-Fusion defines three scalar scores:

1. First-level detail energy: strength of the finest-scale local fluctuations. AI text is usually more uniform at this scale. 2. Multilevel detail energy: total detail energy across all scales, capturing overall multi-scale fluctuation intensity. 3. Window-energy variability: the signal is split into windows; variation in per-window energy captures unevenness of local fluctuations.

These scores are independent and can be used alone or combined.

Calibration-Guided Voting Fusion

Beyond single-score detection, DWT-Fusion introduces "Calibration-Guided Voting Fusion."

The idea: compute scores under multiple wavelet configurations (different wavelet families, different decomposition depths), each casting a "vote," then weight the votes with calibration weights.

This resembles ensemble learning — combining weak detectors into a strong one. Calibration weights ensure better-performing configurations count more.

Experimental Results

The authors tested on three datasets:

HC3: Chinese AI-text detection, relatively easy
M4: multi-generator, multi-domain, medium difficulty
MAGE: multi-generator, multi-domain, multilingual, hardest

Best single-score results:

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

After calibration-guided voting fusion:

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

Near-perfect on HC3 (0.99+), with significant gains on M4 and MAGE. Notably, these results are training-free — no labeled data, no trained classifier, just an open-source proxy model.

Comparison with a DFT Baseline

A key ablation: replace the DWT with the Discrete Fourier Transform (DFT) and see how much performance drops.

DFT gives frequency-domain information only, no time-domain localization. Result: the DFT baseline underperforms DWT-Fusion on all datasets. This directly proves the importance of "local + multi-scale" information — it's not signal processing per se that helps, but the specific joint time-frequency structure of wavelets.

Engineering Significance

1. Zero training cost: no labeled data, no classifier training, just an open-source proxy model. Extremely cheap to deploy. 2. Interpretability: all three scores have clear physical meaning (local fluctuation, multi-scale fluctuation, window variability) — not black boxes. You can analyze why a passage was flagged. 3. Extensibility: wavelet family and decomposition depth are configurable per scenario.

An Honest Assessment

Limitations of this paper:

AUROC of only 0.7471 on MAGE. In deployment this implies a fairly high false-positive rate — insufficient for high-precision scenarios like academic integrity checks.
Proxy-model dependency: the method relies on an open-source model's token probabilities. If text was generated by GPT-4, GPT-2 probabilities may not discriminate well. The paper runs proxy-model sensitivity analysis but does not test GPT-4-generated text.
Adversarial robustness unverified: an adversary who knows the detector uses wavelets could deliberately generate text with "human-like multi-scale fluctuations." Not tested.
Gap to supervised SOTA: the best supervised detectors (RADAR, Ghostbuster) reach AUROC 0.85+ on M4; DWT-Fusion's 0.8477 approaches but does not exceed them. Whether the training-free advantage outweighs the performance gap depends on the deployment scenario.

One-Sentence Summary

DWT-Fusion treats token probabilities as a one-dimensional signal and uses wavelet transforms to capture local and multi-scale probability fluctuations — no training, no labels, just an open-source model — achieving detection performance close to supervised methods.

---

Paper link: https://arxiv.org/abs/2607.22026 HTML version: https://arxiv.org/html/2607.22026v1 Open-source code: not yet available