pith. sign in

arxiv: 2605.28355 · v1 · pith:GVMUZHL6new · submitted 2026-05-27 · 💻 cs.LG

Detecting Diffusion-Generated Time Series Under Generator Shift

Zhi Wen Soi , Aditya Shankar , Gert Lek , Abele M\u{a}lan , Daniel Neider , Jian-Jia Chen , Lydia Chen This is my paper

Pith reviewed 2026-06-29 13:42 UTC · model grok-4.3

classification 💻 cs.LG
keywords diffusion modelstime series detectiongenerator shiftblack-box detectionwhite-box detectionsynthetic datamachine learning
0
0 comments X

The pith

Black-box classifiers detect diffusion-generated time series more reliably than reconstruction methods when the generator is unknown.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that white-box reconstruction detectors, which work in the image domain because of near-universal generators, fail for time series under generator shift. In contrast, a simple off-the-shelf black-box classifier achieves strong performance across shifts, with an average F1 score of 79.2. This indicates that time series detection cannot be treated as a direct transfer of image detection techniques. The work provides the first systematic comparison and identifies open directions for future methods.

Core claim

Reconstruction-based white-box detection succeeds in-distribution but collapses under generator shift for diffusion-generated time series because no analogous near-universal reconstruction prior exists, unlike in images. A black-box detector using an off-the-shelf classifier on the raw signal instead reaches an average F1 of 79.2, a 22.1 percent relative improvement, and a TPR at 1 percent FPR of 57.2, establishing that the problem requires distinct approaches.

What carries the argument

The black-box detector, an off-the-shelf classifier applied directly to raw time series signals, which operates without access to the generator.

If this is right

  • Black-box classification offers a practical route for detection when the diffusion generator is unknown or proprietary.
  • Reconstruction priors that succeed for images do not transfer directly to time series data.
  • Detection performance remains measurable even at low false-positive rates, with TPR reaching 57.2 at 1 percent FPR.
  • The gap between white-box and black-box methods points to the need for modality-specific detector designs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar black-box advantages may appear in other data modalities that lack strong generic reconstruction priors.
  • Practical systems could combine black-box detectors with minimal assumptions about the generative process.
  • The observed robustness under shift suggests testing the same classifier on non-diffusion generative models for time series.

Load-bearing premise

The claim that reconstruction-based detection fails specifically because time series lack a near-universal generator prior, rather than due to other experimental choices such as datasets or shift severity.

What would settle it

A controlled test in which a reconstruction detector achieves high accuracy under generator shift when provided with a sufficiently general time series generator would falsify the central explanation for the performance gap.

Figures

Figures reproduced from arXiv: 2605.28355 by Abele M\u{a}lan, Aditya Shankar, Daniel Neider, Gert Lek, Jian-Jia Chen, Lydia Chen, Zhi Wen Soi.

Figure 2
Figure 2. Figure 2: Cross-generator detection. ○1 Each generator 𝐺𝑖 ∈ G = {𝐺1,𝐺2, . . . ,𝐺𝑁 } generates a synthetic sample 𝑥𝐺𝑖 . ○2 The detector D receives a real sample 𝑥𝑅 along with synthetic samples and ○3 outputs a binary prediction (real or synthetic) for each. to produce a reconstructed sample 𝑥ˆ0 = R𝐺∗ (𝑥0), where R𝐺∗ de￾notes the full DDIM inversion and denoising process under 𝐺 ∗ . We then define the reconstruction e… view at source ↗
Figure 4
Figure 4. Figure 4: Black-box classifier-based detection. The classifier [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: The reconstruction model has learned 𝑝𝐺∗ specific to the reference generator 𝐺 ∗ , so the reconstruction error effectively measures distance to 𝑝𝐺∗ . Samples from an unseen generator 𝐺𝑖 are off-distribution for 𝐺 ∗ in model-specific ways, and real samples are likewise off-distribution. The two distances 𝑑𝐺∗ (𝑥𝐺𝑖 ) and 𝑑𝐺∗ (𝑥𝑅) are of comparable magnitude, so their reconstruction errors overlap and the dete… view at source ↗
Figure 5
Figure 5. Figure 5: Aggregated detectability versus aggregated quality for each generator across ten datasets at different sequence [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original abstract

The boundary between real and diffusion-generated time series is becoming increasingly difficult to draw, yet detection in this domain remains underexplored, especially when the generator is unknown. We compare white-box detection, which requires access to the generator, against black-box detection, which operates on the raw signal alone. The white-box approach, a reconstruction-based detector adapted from the image domain, works well in in-distribution but breaks down under generator shift: reconstruction-based detection in images succeeds because large generic generators provide a near-universal reconstruction prior, and no analogous generator exists for time series. In contrast, a simple off-the-shelf classifier used as a black-box detector performs remarkably well, achieving an average F1 of 79.2, a 22.1% relative improvement over the white-box approach, and a TPR@1%FPR of 57.2. Diffusion-generated time series detection is therefore not a direct transfer of the image domain problem. This work provides the first systematic exploration of white-box and black-box detection for diffusion-generated time series. We close by identifying several open and promising directions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper compares white-box (reconstruction-based, adapted from images) and black-box (off-the-shelf classifier) detection of diffusion-generated time series under generator shift. It reports that the white-box detector succeeds in-distribution but fails under shift because 'large generic generators provide a near-universal reconstruction prior' in images but 'no analogous generator exists for time series.' In contrast, the black-box approach achieves average F1 of 79.2 (22.1% relative improvement over white-box), TPR@1%FPR of 57.2, and concludes that time-series detection is therefore not a direct transfer of the image-domain problem. The work positions itself as the first systematic exploration and identifies open directions.

Significance. If the reported performance gap and its attribution hold after controls, the result would usefully demonstrate that generator-shift robustness in detection does not transfer directly from images to time series, supplying concrete metrics and motivating domain-specific methods. The empirical framing (rather than a parameter-free derivation) makes the strength conditional on the quality of the experimental design.

major comments (2)
  1. [Abstract] Abstract: the central explanatory claim—that reconstruction-based detection 'breaks down under generator shift' specifically because 'no analogous generator exists for time series'—is invoked to support the conclusion that detection 'is therefore not a direct transfer.' No ablations, controls, or comparisons are described that isolate this factor from alternatives such as the particular reconstruction architecture chosen, the magnitude of generator shift in the evaluated datasets, or time-series-specific data properties. This attribution is load-bearing for the non-transfer conclusion yet remains untested in the reported evidence.
  2. [Abstract] Abstract: the performance numbers (F1 79.2, 22.1% relative improvement, TPR@1%FPR 57.2) are presented without accompanying dataset details, number of runs, error bars, statistical tests, or description of the 'off-the-shelf classifier' and white-box baseline architectures. These omissions make it impossible to assess whether the reported gap is robust or sensitive to post-hoc choices.
minor comments (1)
  1. [Abstract] The abstract supplies concrete metrics but supplies no derivation, error bars, dataset details, or statistical tests.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of experimental rigor and claim attribution that we address point-by-point below. We propose targeted revisions to clarify the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central explanatory claim—that reconstruction-based detection 'breaks down under generator shift' specifically because 'no analogous generator exists for time series'—is invoked to support the conclusion that detection 'is therefore not a direct transfer.' No ablations, controls, or comparisons are described that isolate this factor from alternatives such as the particular reconstruction architecture chosen, the magnitude of generator shift in the evaluated datasets, or time-series-specific data properties. This attribution is load-bearing for the non-transfer conclusion yet remains untested in the reported evidence.

    Authors: We agree that the manuscript does not include ablations or controls that isolate the 'no analogous generator' explanation from confounding factors such as architecture choice or the specific magnitude of generator shift. The attribution draws from the performance gap observed in our experiments together with the documented existence of universal reconstruction priors in the image domain. We will revise the abstract to present this as a domain-motivated hypothesis rather than a definitively isolated causal factor, and we will add a dedicated discussion subsection enumerating alternative explanations and their plausibility given the current evidence. revision: yes

  2. Referee: [Abstract] Abstract: the performance numbers (F1 79.2, 22.1% relative improvement, TPR@1%FPR 57.2) are presented without accompanying dataset details, number of runs, error bars, statistical tests, or description of the 'off-the-shelf classifier' and white-box baseline architectures. These omissions make it impossible to assess whether the reported gap is robust or sensitive to post-hoc choices.

    Authors: The abstract is a concise summary; the full experimental protocol—including the exact datasets, the specific classifier and reconstruction architectures, number of runs, and evaluation protocol—is described in the Experiments and Implementation Details sections. We will add error bars, report the number of independent runs, and include basic statistical comparisons (e.g., paired t-tests where appropriate) to the results tables and figures. If the abstract length permits, we will also insert a brief parenthetical reference to the experimental section for the key metrics. revision: partial

Circularity Check

0 steps flagged

No circularity; empirical measurements only

full rationale

The paper reports measured performance numbers (average F1 of 79.2, 22.1% relative improvement, TPR@1%FPR of 57.2) from classifier experiments on time-series data. These are direct empirical outcomes, not quantities derived from equations or parameters that are then re-labeled as predictions. The interpretive claim about absence of a 'near-universal reconstruction prior' for time series is presented as background motivation rather than a load-bearing derivation; it does not reduce any reported result to a fitted input or self-citation chain. No equations, fitted parameters, or self-citation steps are invoked in the provided text to support the central empirical comparison.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no information on free parameters, axioms, or invented entities; all arrays are therefore empty.

pith-pipeline@v0.9.1-grok · 5740 in / 1205 out tokens · 41398 ms · 2026-06-29T13:42:36.922881+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    Juan Miguel Lopez Alcaraz and Nils Strodthoff. 2023. Diffusion-based Time Series Imputation and Forecasting with Structured State Space Models.Trans. Mach. Learn. Res.2023 (2023)

    2023

  2. [2]

    Chronos-2: From Univariate to Universal Forecasting

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael B...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2024. A decoder-only foundation model for time-series forecasting. InICML. PMLR / OpenReview.net, Vienna, Austria, 10148–10167

    2024

  4. [4]

    Schmidt, Jonathan Weber, Geoffrey I

    Hassan Ismail Fawaz, Benjamin Lucas, Germain Forestier, Charlotte Pelletier, Daniel F. Schmidt, Jonathan Weber, Geoffrey I. Webb, Lhassane Idoumghar, Pierre- Alain Muller, and François Petitjean. 2020. InceptionTime: Finding AlexNet for time series classification.Data Min. Knowl. Discov.34, 6 (2020), 1936–1962

    2020

  5. [5]

    Seyed Navid Mohammadi Foumani, Chang Wei Tan, and Mahsa Salehi. 2021. Disjoint-CNN for Multivariate Time Series Classification. InICDM (Workshops). IEEE, Auckland, New Zealand, 760–769

    2021

  6. [6]

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. 2024. MOMENT: A Family of Open Time-series Foundation Models. InICML. PMLR / OpenReview.net, Vienna, Austria, 16115–16152

    2024

  7. [7]

    Xu Guo and Yiqiang Chen. 2024. Generative AI for Synthetic Data Generation: Methods, Challenges and the Future.CoRRabs/2403.04190 (2024)

    work page arXiv 2024

  8. [8]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InNeurIPS, Vol. 33. Curran Associates, Inc., Virtual Event, 6840–6851

    2020

  9. [9]

    Junji Hou, Junzhou Zhao, Shuo Zhang, and Pinghui Wang. 2026. A Theoretical Analysis of Detecting Large Model-Generated Time Series. InAAAI. AAAI Press, Singapore, 21788–21796

    2026

  10. [10]

    Christian Internò, Robert Geirhos, Markus Olhofer, Sunny Liu, Barbara Hammer, and David A. Klindt. 2025. AI-Generated Video Detection via Perceptual Straight- ening. InNeurIPS, Vol. 38. Curran Associates, Inc., San Diego, USA, 20672–20705

    2025

  11. [11]

    Ali Ismail-Fawaz, Maxime Devanne, Stefano Berretti, Jonathan Weber, and Ger- main Forestier. 2023. LITE: Light Inception with boosTing tEchniques for Time Series Classification. InDSAA. IEEE, Thessaloniki, Greece, 1–10

    2023

  12. [12]

    Paul Jeha, Michael Bohlke-Schneider, Pedro Mercado, Shubham Kapoor, Rajbir- Singh Nirwan, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. 2022. PSA-GAN: Progressive Self Attention GANs for Synthetic Time Series. InICLR. OpenReview.net, Virtual Event

    2022

  13. [13]

    Marcel Kollovieh, Abdul Fatir Ansari, Michael Bohlke-Schneider, Jasper Zschieg- ner, Hao Wang, and Yuyang Wang. 2023. Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series Forecasting. InNeurIPS, Vol. 36. Curran Associates, Inc., New Orleans, USA, 28341–28364

    2023

  14. [14]

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2021. DiffWave: A Versatile Diffusion Model for Audio Synthesis. InICLR. OpenRe- view.net, Virtual Event

    2021

  15. [15]

    Hui Li, Yunpeng Cui, Shuo Wang, Juan Liu, Jinyuan Qin, and Yilin Yang. 2020. Multivariate Financial Time-Series Prediction With Certified Robustness.IEEE Access8 (2020), 109133–109143

    2020

  16. [16]

    Liu Sheng, Kensaku Kawamoto, and Samir E

    Mohammad Amin Morid, Olivia R. Liu Sheng, Kensaku Kawamoto, and Samir E. Abdelrahman. 2020. Learning hidden patterns from patient multivariate time series data using convolutional neural networks: A case study of healthcare cost prediction.J. Biomed. Informatics111 (2020), 103565

    2020

  17. [17]

    Hao Ni, Lukasz Szpruch, Magnus Wiese, Shujian Liao, and Baoren Xiao. 2020. Con- ditional Sig-Wasserstein GANs for Time Series Generation.CoRRabs/2006.05421 (2020)

    work page arXiv 2020

  18. [18]

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. 2023. Towards Universal Fake Image Detectors that Generalize Across Generative Models. InCVPR. IEEE, Vancouver, Canada, 24480–24489

    2023

  19. [19]

    Jonas Ricker, Denis Lukovnikov, and Asja Fischer. 2024. AEROBLADE: Training- Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error. InCVPR. IEEE, Seattle, USA, 9130–9140

    2024

  20. [20]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. InCVPR. IEEE, New Orleans, USA, 10674–10685

    2022

  21. [21]

    Aditya Shankar, Lydia Yiyu Chen, Arie van Deursen, and Rihan Hai. 2025. WaveStitch: Flexible and Fast Conditional Time Series Generation With Dif- fusion Models.Proc. ACM Manag. Data3, 6 (2025), 1–25

    2025

  22. [22]

    Galjaard, Huijuan Wang, and Lydia Y

    Zhi Wen Soi, Chaoyi Zhu, Fouad Abiad, Aditya Shankar, Jeroen M. Galjaard, Huijuan Wang, and Lydia Y. Chen. 2025. TimeWak: Temporal Chained-Hashing Watermark for Time Series Data. InNeurIPS, Vol. 38. Curran Associates, Inc., San Diego, USA, 137917–137958

    2025

  23. [23]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising Diffusion Implicit Models. InICLR. OpenReview.net, Virtual Event

    2021

  24. [24]

    Namjoon Suh, Yuning Yang, Din-Yin Hsieh, Qitong Luan, Shirong Xu, Shixiang Zhu, and Guang Cheng. 2024. TimeAutoDiff: Combining Autoencoder and Diffusion model for time series tabular data synthesizing.CoRRabs/2406.16028 (2024)

    work page arXiv 2024

  25. [25]

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. 2020. CNN-Generated Images Are Surprisingly Easy to Spot... for Now. In CVPR. Computer Vision Foundation / IEEE, Seattle, USA, 8692–8701

    2020

  26. [26]

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. 2023. DIRE for Diffusion-Generated Image Detection. In ICCV. IEEE, Paris, France, 22388–22398

    2023

  27. [27]

    Zhiguang Wang, Weizhong Yan, and Tim Oates. 2017. Time series classification from scratch with deep neural networks: A strong baseline. InIJCNN. IEEE, Anchorage, USA, 1578–1585

    2017

  28. [28]

    Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Weidi Xie. 2025. A Sanity Check for AI-generated Image Detection. InICLR. OpenReview.net, Singapore

    2025

  29. [29]

    Jinsung Yoon, Daniel Jarrett, and Mihaela van der Schaar. 2019. Time-series Generative Adversarial Networks. InNeurIPS, Vol. 32. Curran Associates, Inc., Vancouver, Canada, 5509–5519

    2019

  30. [30]

    Xinyu Yuan and Yan Qiao. 2024. Diffusion-TS: Interpretable Diffusion for General Time Series Generation. InICLR. OpenReview.net, Vienna, Austria

    2024

  31. [31]

    Bendong Zhao, Huanzhang Lu, Shangfeng Chen, Junliang Liu, and Dongya Wu

  32. [32]

    Convolutional neural networks for time series classification.J. Syst. Eng. Electron.28, 1 (2017), 162–169. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Soi et al. Table 3: All metrics are higher the better, with TPR measured at a 1% FPR. Values in gray are ID, while values in black are OOD. The best average OOD scores are highlighted in green ....

    2017