DWT-Fusion: Detecting AI-Generated Text with Wavelet Transforms on Token Probability Signals

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

The Scenario

You receive a student essay and want to know whether the student wrote it or whether it was generated by ChatGPT. You have access to an open-source language model (e.g., GPT-2 or LLaMA), so you can compute the log-probability of every token in the text. But how do you turn those probabilities into a verdict?

The most intuitive approach: compute the average log-probability of the whole text. AI-generated text tends to have higher probability (more "fluent"), while human writing tends to have lower probability (less "fluent").

But this has a fatal flaw: a skilled human writer's fluent prose also scores high, and AI text deliberately made to imitate human "roughness" may score low. The global average hides local differences.

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

Turning Text into a Signal

The first step is converting the text into a one-dimensional signal. A proxy language model (e.g., GPT-2 Medium) computes the conditional log-probability of each token. This sequence of numbers is the signal—like an audio waveform.

Then the method applies a Discrete Wavelet Transform (DWT) to this signal.

Why Wavelets?

The wavelet transform is a classic signal-processing tool: it decomposes a signal into "details" at different scales. Global statistics (mean, variance) capture overall brightness of a photo but no detail; Fourier analysis tells you which frequencies exist but not where. Wavelets tell you both what frequency and where—joint time and frequency information.

DWT-Fusion uses wavelets because differences between AI and human text may be local and multi-scale rather than global:

• Local differences: AI text may be extremely "fluent" (high, stable probability) in some passages but deliberately "rough" in others to mimic humans. A global average cancels these out.
• Multi-scale differences: AI probability fluctuations may be small at short scales (a few tokens) but show regular patterns at long scales (tens of tokens). Human text fluctuates differently.

The DWT decomposes the signal into approximation coefficients (low-frequency, global trend) and detail coefficients (high-frequency, local variation) across multiple levels—exactly the structure needed to capture both kinds of differences.

Three Wavelet-Domain Scores

DWT-Fusion defines three scalar scores:

1. First-level detail energy — intensity of the finest-scale local fluctuations. AI text typically fluctuates less at this scale (more uniform). 2. Multilevel detail energy — total detail energy across all scales, capturing overall multi-scale fluctuation intensity. 3. Window-energy variability — splitting the signal into windows and measuring how per-window energy varies, capturing unevenness of local fluctuations.

Each score works independently as a detection signal, and they can also be combined.

Calibration-Guided Voting Fusion

Beyond single-score detection, DWT-Fusion introduces Calibration-Guided Voting Fusion: multiple wavelet configurations (different wavelet families, different decomposition depths) each produce a score and cast a "vote," weighted by calibration weights so better-performing configurations count more. It's like ensemble learning—many weak detectors combined into a strong one.

Experimental Results

The authors evaluate 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 classifier training, just an open-source proxy model.

Comparison with a DFT Baseline

A key ablation replaces DWT with the Discrete Fourier Transform (DFT), which provides frequency but no time-domain (positional) information. The DFT baseline underperforms DWT-Fusion on all datasets. This directly demonstrates the importance of local + multi-scale information—it's not signal processing per se that helps, but the wavelet's specific joint time-frequency structure.

Engineering Significance

1. Zero training cost. No labeled data or classifier training; just an open-source proxy model. Deployment cost is minimal. 2. Interpretability. The three scores have clear physical meaning (local fluctuation, multi-scale fluctuation, window variability)—not a black box. You can analyze why a passage was flagged. 3. Configurability. Wavelet family and decomposition depth are tunable per scenario.

An Honest Assessment

Limitations worth noting:

• AUROC of only 0.7471 on MAGE. In practice this implies a fairly high false-positive rate—insufficient for high-precision scenarios like academic integrity enforcement.
• Proxy-model dependence. The method relies on token probabilities from an open-source model. Text generated by GPT-4 may not be well distinguished using GPT-2 probabilities. The paper includes a proxy-model sensitivity analysis but no tests on GPT-4-generated text.
• Adversarial robustness unverified. An adversary aware of the wavelet-based detector could generate text whose multi-scale fluctuations mimic humans. Not tested.
• Gap to supervised SOTA. Best supervised detectors (RADAR, Ghostbuster) reach AUROC 0.85+ on M4; DWT-Fusion's 0.8477 approaches but does not exceed them. Whether training-free convenience outweighs the performance gap depends on the deployment context.

One-Line 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—approaching supervised detectors' performance in AI-text detection.

---

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