Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Per-instance hypernetwork optimization over a frozen diffusion model generates targeted, realistic failure cases for deep learning systems without curated conditioning data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-03 08:57 UTC pith:JOV5W5XO

load-bearing objection Solid method, inflated headline: HyNeA works and deserves review, but the driving experiment contradicts the dataset-free claim. the 2 major comments →

arxiv 2601.15041 v2 pith:JOV5W5XO submitted 2026-01-21 cs.LG cs.SE

HyperNet-Adaptation for Diffusion-Based Test Case Generation

classification cs.LG cs.SE
keywords diffusion modelshypernetworktest case generationdeep learning testingdataset-free controllabilityobject detectionimage classificationgenerative testing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces HyNeA, a method for generating test images that make a deep learning system fail in a chosen, targeted way. Instead of training a conditioning network on curated examples of failures or fine-tuning the generator, HyNeA optimizes a small auxiliary hypernetwork separately for each test case, backpropagating the target system's prediction loss through a frozen diffusion model. The paper claims this makes controllable generation dataset-free and general: across image classification, multi-label attribute prediction, and object detection, all generated test cases flipped the target output while staying visually close to the original images. The practical payoff is that testers can search for realistic failures without needing failure-labeled data, and at a fraction of the SUT evaluations used by search-based generators.

Core claim

The central claim is that controllability in diffusion-based test generation can be achieved without conditioning datasets or architecture-specific prompts, by inverting the ControlNet idea: instead of learning a fixed mapping from control signals to images, HyNeA starts from an image and adapts a hypernetwork so that the image produced by the frozen diffusion backbone yields a desired prediction from the system under test. The adaptation is instance-level: the hypernetwork weights are optimized for each generated test case, with the loss computed on the final denoised image rather than per denoising step, which the paper argues gives more informative gradients. Empirically, HyNeA reports pe

What carries the argument

The load-bearing component is a ControlNet-style auxiliary hypernetwork attached to a frozen latent diffusion model, together with a trainable control projector that maps arbitrary SUT outputs (class logits, attribute vectors, detection tensors) into a spatial conditioning signal. Zero-initialized adapter layers let the hypernetwork modulate the denoising process without retraining the backbone. During test generation, AdamW with a OneCycle schedule optimizes the hypernetwork weights per instance by backpropagating a sum of a visual-fidelity loss (Frobenius distance to the original image) and a behavior-steering loss (cross-entropy or binary cross-entropy against a chosen failure target) thr

Load-bearing premise

The method assumes that gradients of the system-under-test's loss, backpropagated through the entire diffusion denoising chain into the hypernetwork, are informative enough that a fixed short optimization schedule can steer generation toward a target failure without destroying image coherence.

What would settle it

Run HyNeA with its recommended learning rate on a new diffusion backbone and a new SUT whose output space is not smoothly encoded in the latent space (e.g., a detector with rare classes); if the fraction of generated images that achieve the targeted misclassification while being judged label-preserving falls far below the roughly 0.79 human label-preservation rate reported on the classification task, the general claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, testers can generate targeted, realistic failures for any SUT whose output can be turned into a spatial control signal, with no failure-labeled dataset.
  • The reported efficiency (on the order of 6–30 SUT evaluations per test case, versus hundreds or thousands for search-based methods) means gradient-driven diffusion adaptation could scale to larger test suites.
  • Because the diffusion backbone is frozen, the method inherits the generator's realism and can be re-targeted by swapping only the loss and the SUT, suggesting reuse across systems under test.
  • The near-zero escape ratios reported imply that generated test cases hit the intended failure mode rather than degrading images into generic misclassifications, which matters for debugging.
  • The paper's comparison suggests that GAN-based and noise-perturbation baselines are more likely to push latents out of distribution, while HyNeA's semantic steering keeps generated images in-distribution.

Where Pith is reading between the lines

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

  • HyNeA's control projector is described generically, so the same mechanism should extend to SUTs with structured outputs such as segmentation maps, keypoints, or depth; testing that extension on a segmentation model would be a direct check of the method's generality.
  • The equal weighting of fidelity and behavior losses is a choice, not a requirement; tuning their relative weight could let testers trade between near-boundary failures and deeper semantic shifts, or produce hard vs. easy test cases.
  • The steady degradation with increasing learning rate suggests a predictable reliability curve: a small learning-rate sweep on a few seed images may be enough to calibrate HyNeA for a new task, making the per-task tuning cost modest.
  • Because the method only needs gradients from the SUT and a frozen generative model, it could plausibly be applied to non-image modalities such as audio or time series whenever a latent diffusion model exists; this is an extrapolation the paper does not claim.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper presents HyNeA, a generative test-case generation method for deep learning vision systems. HyNeA builds on ControlNet-style architectures by introducing a trainable hypernetwork that modulates a frozen pretrained diffusion model. Rather than training the hypernetwork on curated image-control pairs, HyNeA optimizes its weights per test instance by backpropagating a behavior-driven loss through the latent diffusion process, thereby steering the generated image toward a user-defined failure of the system under test. The authors evaluate HyNeA on ImageNet classification, CelebA binary attribute classification, and KITTI-style driving object detection, comparing against Mimicry (StyleGAN-based) and GIFTbench (diffusion-based). They report that HyNeA achieves misclassification rates of 1.0 across tasks, lower escape ratios, higher structural similarity, and improved human-rated realism, at lower computational budgets than the baselines.

Significance. If the method and results hold, HyNeA would be a meaningful advance in DL testing: it offers a way to generate targeted, realistic failure cases without a pre-collected set of failure examples, and it shows that per-instance hypernetwork adaptation is a viable control mechanism for diffusion models. The paper includes a replication package, a three-task evaluation, and a useful theoretical appendix explaining why GIFTbench's latent noise perturbation tends to produce out-of-distribution samples. The main empirical claims of realism and diversity are supported by human evaluation (albeit on ImageNet only) and by image-quality metrics.

major comments (2)
  1. [§4.4.3 and abstract] The abstract and Section 1 claim that HyNeA works 'without relying on architecture-specific conditioning mechanisms or dataset-driven adaptations such as fine-tuning.' The Driving evaluation in §4.4.3 directly contradicts this: the diffusion model is 'a Stable Diffusion 1.5 model fine-tuned on real-world road scenes [5]' and generation is 'guided using a segmentation-based ControlNet [58] trained on driving-scene segmentations, with KITTI-derived segmentation maps serving as control signals.' These are precisely the adaptations the claim excludes. As a result, the Driving results (Table 1, misclassification rate 1.00; MS-SSIM 0.95) do not demonstrate dataset-free control; they demonstrate control over a heavily domain-adapted pipeline. Please either remove the dataset-free claim for this experiment, or replace the base model with an unmodified one and re-run. This is load-bearing because
  2. [§3.4, §4.6.1, Table 1] The RQ1 effectiveness metrics are directly aligned with the optimization objectives in Eq. (3) and the loss definitions in §3.4: misclassification rate, escape ratio, and confidence reduction are all computed from exactly the CE/BCE terms being minimized. For example, Eq. (10)–(12) measure whether argmax flips, which is the same signal used in the behavior-steering loss. Consequently, reporting 'misclassification rate = 1.0' (Table 1) is close to a statement that the optimizer converged, not that the generated test cases are more effective in a meaningful comparative sense. The paper should either (a) restrict the RQ1 claim to the non-circular metrics (MS-SSIM, LPIPS, diversity, trace difference, human realism) and relegate the loss-aligned metrics to a sanity check, or (b) justify why the comparison against baselines with different objectives is fair. As written, the RQ1 conclusion over
minor comments (5)
  1. [§4.5, Table 1] The dataset identifiers in Table 1 and §4.5 appear as damaged glyphs: '⊷ImageNet', '/userCelebA', '/carDriving'. Please replace with standard text.
  2. [Table 1] Report standard deviations for misclassification rate and escape ratio where these are not constants; currently only MS-SSIM and LPIPS include uncertainty. Also clarify the column header 'Diversity T' (does it represent trace of the covariance, a distinct measure?).
  3. [Abstract / §1] The claim '20–100% more relevant failure-inducing test cases' does not appear to be tied to any table or calculation in §4. Please either provide the exact derivation or adjust the claim to match the reported metrics.
  4. [§4.6.3] The human evaluation is conducted only on ImageNet, as disclosed in §4.4.3, but the Conclusion states 'Human evaluators rate HyNeA's test cases as more realistic' without this scope restriction. Please qualify the statement to avoid overgeneralization.
  5. [Table 2] The 'Mimicry(Diffusion)' rows are extrapolations, not measured baselines. The tilde is a good indicator, but the text in §4.6.2 should explicitly state that these are estimates and are not used in any statistical comparison.

Circularity Check

1 steps flagged

RQ1 effectiveness metrics are the optimized loss; realism and diversity results are independent. The dataset-free claim is undermined by the Driving setup but that is a consistency issue, not circularity.

specific steps
  1. fitted input called prediction [Section 3.4 (Eq. 5) vs. Section 4.2.1 (Eqs. 10, 13, 16); discussed in Section 4.6.1]
    "HyNeA achieves a perfect value of 0, indicating that all generated test cases trigger their targeted misbehavior, which is an effect of the loss terms, steering towards specific targets."

    Per instance, HyNeA fits HyperNet weights by minimizing L_C = ω_dF + ω_CE (and the BCE/CEm analogues), where ω_CE pushes the SUT's prediction toward a chosen target class. The RQ1 effectiveness metrics are direct functionals of the same objective: misclassification rate counts an argmax change, escape ratio checks that the new top class is one of the intended targets, and confidence reduction averages the drop in the original-class logit—exactly what cross-entropy minimization produces. Thus mr=1.00 and escape=0.00 restate convergence of the fit rather than an independent discovery of failures. The paper's own 'effect of the loss terms' acknowledges this. MS-SSIM, LPIPS, embedding diversity, and human realism are not directly optimized and provide independent evidence, so the circularity i

full rationale

The only identifiable circularity is the alignment between the RQ1 effectiveness metrics and the per-instance optimization objective. Because the HyperNet is fitted by minimizing the same cross-entropy/BCE loss that the metrics threshold, the perfect misclassification and escape numbers are partly expected; the paper itself states this is 'an effect of the loss terms.' I score this as moderate and partial, not total, because (i) the comparison with Mimicry and GIFTbench remains informative—all methods receive the same loss terms, and HyNeA reaches them more reliably—and (ii) the realism/semantic-validity claims rest on independent measurements (MS-SSIM, LPIPS, embedding diversity, human annotation) that are not directly optimized. Separately, the 'dataset-free' headline claim is internally inconsistent with the Driving setup in Section 4.4.3, which uses a road-scene fine-tuned Stable Diffusion and a KITTI-trained segmentation ControlNet; this is a validity/consistency problem, not a circular derivation, so it does not by itself raise the circularity score. No self-citation is load-bearing: citations to the authors' own Mimicry/GIFTbench/Detect work are baselines or related-work comparisons, not justifications of the method's correctness.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The HyperNet is an auxiliary neural network adapted from ControlNet, not a new postulational entity. The main free parameters are learning rates and optimization steps, tuned per task. The key domain assumptions are differentiability of the SUT and the diffusion model, and the suitability of the control projector.

free parameters (5)
  • learning rate minimum (lr_min) = ImageNet: 1e-6; CelebA: 1e-5; Driving: 1e-4 (selected from sweeps)
    RQ4 sweeps learning rates and chooses the schedule that balances runtime, realism, and MS-SSIM. These values are task-specific and directly influence success.
  • learning rate maximum (lr_max) = 100 * lr_min for ImageNet; ratios vary for CelebA and Driving
    The OneCycle schedule requires lr_max; the paper fixes lr_max = 100 * lr_min for ImageNet and specifies alternations for other tasks.
  • optimization steps = 2500 (with early stopping)
    The number of gradient updates per test case; chosen to produce small incremental changes without destabilizing the generator (§3.3).
  • objective coefficients alpha_i = 1 for all losses
    Equal weighting of visual fidelity and behavior steering (§3).
  • top-5 detection restriction = 5
    For object detection, only the top-5 detections are considered to avoid dilution of the loss (§3.1).
axioms (5)
  • domain assumption The SUT is differentiable and accessible in a white-box manner.
    The method backpropagates through the SUT's outputs to compute gradients for the HyperNet. This is stated in the Discussion: 'HyNeA ... employs a white-box approach.'
  • domain assumption The diffusion model's denoising process is differentiable and supports gradient backpropagation.
    Section 4.5 states 'we backpropagate through the diffusion model to generate failure-inducing perturbations.' This is required for the optimization.
  • domain assumption The control projector (Algorithm 1) can map arbitrary SUT outputs to spatial control signals.
    Section 3.2.1 assumes that classification logits, attribute vectors, and detection tensors can be projected into the latent space with sufficient fidelity.
  • domain assumption Pretrained diffusion models (REPA-E, Stable Diffusion 1.5, LDM-CelebA) reproduce high-quality images and are representative of modern generators.
    The method relies on these backbones; their quality affects the validity of generated test cases (§4.4).
  • domain assumption Human annotations (with attention checks) provide a reliable ground truth for realism and label preservation.
    The human study is the primary evidence for semantic validity; inter-rater agreement is only fair (Fleiss kappa 0.336), but the authors use statistical tests.

pith-pipeline@v1.3.0-alltime-deepseek · 24340 in / 9096 out tokens · 99729 ms · 2026-08-03T08:57:45.992700+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HyperNet-Adaptation for Diffusion-Based Test Case Generation." pith.science (2026). https://pith.science/paper/JOV5W5XO

@misc{pith2026260115041,
  author       = {Pith},
  title        = {Pith review of: HyperNet-Adaptation for Diffusion-Based Test Case Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JOV5W5XO}},
  note         = {Machine review of arXiv:2601.15041}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The increasing deployment of deep learning systems requires systematic evaluation of their reliability in real-world scenarios. Traditional gradient-based adversarial attacks introduce small perturbations that rarely correspond to realistic failures and mainly assess robustness rather than functional behavior. Generative test generation methods offer an alternative but are often limited to simple datasets or constrained input domains. Although diffusion models enable high-fidelity image synthesis, their computational cost and limited controllability restrict their applicability to large-scale testing. We present HyNeA, a generative testing method that enables direct and efficient control over diffusion-based generation. HyNeA provides dataset-free controllability through hypernetworks, allowing targeted manipulation of the generative process without relying on architecture-specific conditioning mechanisms or dataset-driven adaptations such as fine-tuning. HyNeA employs a distinct training strategy that supports instance-level tuning to identify failure-inducing test cases without requiring datasets that explicitly contain examples of similar failures. This approach enables the targeted generation of realistic failure cases at substantially lower computational cost than search-based methods. Experimental results show that HyNeA improves controllability and test diversity compared to existing generative test generators and generalizes to domains where failure-labeled training data is unavailable.

Figures

Figures reproduced from arXiv: 2601.15041 by Andrea Stocco, Oliver Wei{\ss}l, Severin Kacianka, Vincenzo Riccio.

Figure 1
Figure 1. Figure 1: The diffusion and denoising process in a diffusion model. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Simplified ControlNet Architecture (HyperNet + LDM), [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: HyNeA’s component interaction throughout Optimization. Each component and its respective In- and Outputs are described in the following sections. unconstrained generation, while larger values increasingly bias generation toward the specified control. In practice, 𝛾 is realised using zero layers (indicated as 0 in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example Origins and Targets for ë ImageNet. Predicted Origin and Target class are shown on top of the images. Teddy Pizza Embedding Projection (Origin vs Target) Method Mimicry HyNeA GIFTbench Origins Targets [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Embedding comparison for “pizza” and “teddy bear” across methods. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example Origins and Targets for CelebA. Change of predicted attribute presence on top of the images. car traffic light (a) HyNeA car traffic light (b) GIFTbench [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example Origins and Targets for ˆ Driving. The targeted objects are shown above the images, with detected bounding boxes and their corresponding confidence scores displayed. unclear. Although embedding distances do not directly measure semantic validity, larger deviations tend to coincide with images that annotators judged as less realistic or semantically unclear. Samples generated by Mimicry and GIFTbenc… view at source ↗
Figure 9
Figure 9. Figure 9: Correlation between FID and Human anno￾tated Image-Realism. Finally, the Driving task in [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Examples of HyNeA outputs for fooling the SUT with initial class "pizza" across lr-schedules. , Vol. 1, No. 1, Article . Publication date: January 2026 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Runtime–quality trade-off across learning-rate schedules. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Example Origins and Targets for ë ImageNet. Predicted Origin and Target class are shown to the left and right of the images. , Vol. 1, No. 1, Article . Publication date: January 2026 [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Example Origins and Targets for CelebA. Predicted attribute presence is shown in the top left corner of the examples, the difference between Origin and Target is shown in the last column. , Vol. 1, No. 1, Article . Publication date: January 2026 [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Example Origins and Targets for the ˆ Driving Dataset. Predicted Objects are shown with bounding boxes and their confidence. NMS was used to reduce duplicate bounding boxes and a confidence threshold of 0.5 is used to filter further. Prompt P.1 "A photorealistic urban traffic scene with cars, traffic lights, and stop signs, clear skies, daytime, featuring a {class}" Prompt P.2 "blurry, distorted, ugly, lo… view at source ↗
Figure 15
Figure 15. Figure 15: Probability of OOD latents over iterations in [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: The Anderson–Darling test statistic with a a criticality threshold of [PITH_FULL_IMAGE:figures/full_fig_p032_16.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Generative Testing of Automated Speech Recognition Systems

    cs.CR 2026-07 conditional novelty 6.0

    Phoneme-level latent interpolation in a TTS model yields ~98% black-box ASR failures with higher naturalness than waveform attacks and quality competitive with white-box PGD.

Reference graph

Works this paper leans on

59 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Replication Package

    2026. Replication Package. https://github.com/oliverweissl/SMOO/tree/archive/hynea

  2. [2]

    Shenglin Bao, Nan Jiang, Weijie Zhu, and Pei Zhang. 2024. Generative Model-Based Test Case Generation and Operational Testing for Deep Learning. In2024 5th International Conference on Big Data & Artificial Intelligence & Software Engineering (ICBASE). IEEE, 565–570

  3. [3]

    Luciano Baresi, Davide Yi Xian Hu, Andrea Stocco, and Paolo Tonella. 2025. Efficient Domain Augmentation for Autonomous Driving Testing Using Diffusion Models. In2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 743–743

  4. [4]

    Igor Buzhinsky, Arseny Nerinovsky, and Stavros Tripakis. 2023. Metrics and methods for robustness evaluation of neural networks with generative models.Machine Learning112, 10 (2023), 3977–4012

  5. [5]

    cccccsys. 2024. Guericke Dataset 2. https://huggingface.co/datasets/cccccsys/guericke_dataset_2

  6. [6]

    Xingcheng Chen, Oliver Weissl, and Andrea Stocco. 2026. Feature-Aware Test Generation for Deep Learning Models. arXiv:2601.14081 [cs.SE] https://arxiv.org/abs/2601.14081

  7. [7]

    CompVis Team. 2022. CompVis: ldm-celebahq-256. https://huggingface.co/CompVis/ldm-celebahq-256. Accessed: 2025-11-20

  8. [8]

    Francesco Croce and Matthias Hein. 2020. Minimally distorted adversarial examples with a fast adaptive boundary attack. InInternational conference on machine learning. PMLR, 2196–2205. , Vol. 1, No. 1, Article . Publication date: January 2026. 24 Weißl et al

  9. [9]

    Swaroopa Dola, Rory McDaniel, Matthew B Dwyer, and Mary Lou Soffa. 2024. CIT4DNN: Generating Diverse and Rare Inputs for Neural Networks Using Latent Space Combinatorial Testing. InProceedings of the IEEE/ACM 46th International Conference on Software Engineering. 1–13

  10. [10]

    Alessio Gambi, Marc Mueller, and Gordon Fraser. 2019. Automatically Testing Self-driving Cars with Search-based Procedural Content Generation. InProceedings of the 28th ACM SIGSOFT International Symposium on Software Testing and Analysis(Beijing, China)(ISSTA 2019). ACM, New York, NY, USA, 318–328. doi:10.1145/3293882.3330566

  11. [11]

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. 2013. Vision meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR)(2013)

  12. [12]

    Jianmin Guo, Yu Jiang, Yue Zhao, Quan Chen, and Jiaguang Sun. 2018. Dlfuzz: Differential fuzzing testing of deep learning systems. InProceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering. 739–743

  13. [13]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

  14. [14]

    Jung Im Choi and Qing Tian. 2022. Adversarial attack and defense of yolo detectors in autonomous driving scenarios. In2022 IEEE intelligent vehicles symposium (IV). IEEE, 1011–1017

  15. [15]

    Linfeng Jiang and Peidong Zhan. 2025. FM-YOLOv11: A Lightweight Traffic Sign Detection Network Based on Mixed Attention. InProceedings of the 2025 International Conference on Artificial Intelligence and Computational Intelligence. 433–439

  16. [16]

    2023.Ultralytics YOLO

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. 2023.Ultralytics YOLO. https://github.com/ultralytics/ultralytics

  17. [17]

    Sungmin Kang, Robert Feldt, and Shin Yoo. 2020. Sinvad: Search-based image space navigation for dnn image classifier test input generation. InProceedings of the IEEE/ACM 42nd International Conference on Software Engineering Workshops. 521–528

  18. [18]

    Sungmin Kang, Robert Feldt, and Shin Yoo. 2024. Deceiving Humans and Machines Alike: Search-Based Test Input Generation for DNNs Using Variational Autoencoders.ACM Transactions on Software Engineering Methodologies33 (dec 2024), 103:1–24. Issue 4. doi:10.1145/3635706

  19. [19]

    Tero Karras, Samuli Laine, and Timo Aila. 2019. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4401–4410

  20. [20]

    Diederik P Kingma. 2013. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114(2013)

  21. [21]

    Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. 2018. Adversarial examples in the physical world. InArtificial intelligence safety and security. Chapman and Hall/CRC, 99–112

  22. [22]

    Stefano Carlo Lambertenghi and Andrea Stocco. 2024. Assessing Quality Metrics for Neural Reality Gap Input Mitigation in Autonomous Driving Testing. In2024 IEEE Conference on Software Testing, Verification and Validation (ICST ’24). IEEE, 173–184. doi:10.1109/ICST60714.2024.00024

  23. [23]

    J Richard Landis and Gary G Koch. 1977. The measurement of observer agreement for categorical data.Biometrics (1977), 159–174

  24. [24]

    Xingjian Leng, Jaskirat Singh, Yunzhong Hou, Zhenchang Xing, Saining Xie, and Liang Zheng. 2025. Repa-e: Unlocking vae for end-to-end tuning with latent diffusion transformers.arXiv preprint arXiv:2504.10483(2025)

  25. [25]

    Yue Liu, Lichao Feng, Xingya Wang, and Shiyu Zhang. 2022. DeepBoundary: A Coverage Testing Method of Deep Learning Software based on Decision Boundary Representation. In2022 IEEE 22nd International Conference on Software Quality, Reliability, and Security Companion (QRS-C). IEEE, 166–172

  26. [26]

    Ilya Loshchilov, Frank Hutter, et al. 2017. Fixing weight decay regularization in adam.arXiv preprint arXiv:1711.05101 5, 5 (2017), 5

  27. [27]

    2016.TorchVision: PyTorch’s Computer Vision library

    TorchVision maintainers and contributors. 2016.TorchVision: PyTorch’s Computer Vision library

  28. [28]

    Maryam Maryam, Matteo Biagiola, Andrea Stocco, and Vincenzo Riccio. 2025. Benchmarking Generative AI Models for Deep Learning Test Input Generation. In2025 IEEE Conference on Software Testing, Verification and Validation (ICST). IEEE, 174–185

  29. [29]

    Sondess Missaoui, Simos Gerasimou, and Nicholas Matragkas. 2023. Semantic data augmentation for deep learning testing using generative AI. In2023 38th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, 1694–1698

  30. [30]

    Nusrat Jahan Mozumder, Felipe Toledo, Swaroopa Dola, and Matthew B Dwyer. 2025. RBT4DNN: Requirements-based Testing of Neural Networks.arXiv preprint arXiv:2504.02737(2025)

  31. [31]

    Eliya Nachmani, Robin San Roman, and Lior Wolf. 2021. Non gaussian denoising diffusion models.arXiv preprint arXiv:2106.07582(2021)

  32. [32]

    Neelofar Neelofar and Aldeida Aleti. 2024. Identifying and Explaining Safety-critical Scenarios for Autonomous Vehicles via Key Features.ACM Transactions on Software Engineering and Methodology33, 4 (2024), 1–32

  33. [33]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision. 4195–4205. , Vol. 1, No. 1, Article . Publication date: January 2026. HyperNet-Adaptation for Diffusion-Based Test Case Generation 25

  34. [34]

    Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. 2017. Deepxplore: Automated whitebox testing of deep learning systems, In proceedings of the 26th Symposium on Operating Systems Principles (Shanghai, China).Commun. ACM 62, 11, 1–18. doi:10.1145/3132747.3132785

  35. [35]

    Zipeng Qi, Lichen Bai, Haoyi Xiong, and Zeke Xie. 2024. Not all noises are created equally: Diffusion noise selection and optimization.arXiv preprint arXiv:2407.14041(2024)

  36. [36]

    Vincenzo Riccio, Nargiz Humbatova, Gunel Jahangirova, and Paolo Tonella. 2021. DeepMetis: Augmenting a Deep Learning Test Set to Increase its Mutation Score. InProceedings of the 36th IEEE/ACM International Conference on Automated Software Engineering (ASE ’21). IEEE/ACM

  37. [37]

    Vincenzo Riccio, Gunel Jahangirova, Andrea Stocco, Nargiz Humbatova, Michael Weiss, and Paolo Tonella. 2020. Testing machine learning based systems: a systematic mapping.Empirical Software Engineering25 (2020), 5193–5254

  38. [38]

    Vincenzo Riccio and Paolo Tonella. 2020. Model-based exploration of the frontier of behaviours for deep learning system testing. InProceedings of the 28th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering. 876–888

  39. [39]

    Vincenzo Riccio and Paolo Tonella. 2023. When and Why Test Generators for Deep Learning Produce Invalid Inputs: an Empirical Study. InProceedings of 45th International Conference on Software Engineering (ICSE ’23). ACM, 12 pages

  40. [40]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

  41. [41]

    Axel Sauer, Katja Schwarz, and Andreas Geiger. 2022. Stylegan-xl: Scaling stylegan to large diverse datasets. InACM SIGGRAPH 2022 conference proceedings. 1–10

  42. [42]

    Dario Shariatian, Umut Simsekli, and Alain Oliviero Durmus. [n. d.]. Heavy-Tailed Diffusion with Denoising Levy Probabilistic Models. InThe Thirteenth International Conference on Learning Representations

  43. [43]

    Leslie N Smith and Nicholay Topin. 2019. Super-convergence: Very fast training of neural networks using large learning rates. InArtificial intelligence and machine learning for multi-domain operations applications, Vol. 11006. SPIE, 369–386

  44. [44]

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. 2015. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational conference on machine learning. pmlr, 2256–2265

  45. [45]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502(2020)

  46. [46]

    Yang Song and Stefano Ermon. 2019. Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems32 (2019)

  47. [47]

    Alexander Sorokin and David Forsyth. 2008. Utility data annotation with amazon mechanical turk. In2008 IEEE computer society conference on computer vision and pattern recognition workshops. IEEE, 1–8

  48. [48]

    Yuchi Tian, Kexin Pei, Suman Jana, and Baishakhi Ray. 2018. DeepTest: Automated Testing of Deep-neural-network- driven Autonomous Cars. InProceedings of the 40th International Conference on Software Engineering(Gothenburg, Sweden)(ICSE ’18). ACM, New York, NY, USA, 303–314. doi:10.1145/3180155.3180220

  49. [49]

    Zhiyong Tian and Yunfei Li. 2024. Pedestrian Detection Using Event Cameras and YOLOv8: An Optimized Event Stream to Event Frame Conversion Algorithm. InProceedings of the 2024 9th International Conference on Intelligent Information Processing. 134–140

  50. [50]

    Jente Vandersanden, Sascha Holl, Xingchang Huang, and Gurprit Singh. 2024. Edge-preserving noise for diffusion models. (2024)

  51. [51]

    Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. [n. d.].Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers

  52. [52]

    Zhou Wang, Eero P Simoncelli, and Alan C Bovik. 2003. Multiscale structural similarity for image quality assessment. InThe thrity-seventh asilomar conference on signals, systems & computers, 2003, Vol. 2. Ieee, 1398–1402

  53. [53]

    Oliver Weißl, Amr Abdellatif, Xingcheng Chen, Giorgi Merabishvili, Vincenzo Riccio, Severin Kacianka, and Andrea Stocco. 2025. Targeted Deep Learning System Boundary Testing.ACM Transactions on Software Engineering and Methodology(2025)

  54. [54]

    Ari Wibowo, Bambang Riyanto Trilaksono, Egi Muhammad Idris Hidayat, and Rinaldi Munir. 2023. Object detection in dense and mixed traffic for autonomous vehicles with modified yolo.IEEE Access11 (2023), 134866–134877

  55. [55]

    Frank Wilcoxon. 1945. Individual Comparisons by Ranking Methods.Biometrics Bulletin1, 6 (dec 1945), 80. doi:10. 2307/3001968

  56. [56]

    Jing Yu, Xiaojun Ye, and Qiang Tu. 2022. Traffic sign detection and recognition in multiimages using a fusion model with YOLO and VGG network.IEEE Transactions on Intelligent Transportation Systems23, 9 (2022), 16632–16642

  57. [57]

    Fuyuan Zhang, Sankalan Pal Chowdhury, and Maria Christakis. 2020. Deepsearch: A simple and effective blackbox attack for deep neural networks. InProceedings of the 28th ACM Joint Meeting on European Software Engineering , Vol. 1, No. 1, Article . Publication date: January 2026. 26 Weißl et al. Conference and Symposium on the Foundations of Software Engine...

  58. [58]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF international conference on computer vision. 3836–3847

  59. [59]

    A photorealistic urban traffic scene with cars, traffic lights, and stop signs, clear skies, daytime, featuring a {class}

    Tahereh Zohdinasab, Vincenzo Riccio, Alessio Gambi, and Paolo Tonella. 2021. Deephyperion: exploring the feature space of deep learning-based systems through illumination search. InProceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA ’21). Association for Computing Machinery, 79–90. , Vol. 1, No. 1, Article ....