Pith. sign in

REVIEW 3 major objections 5 minor 41 references

CLOAK: Contrastive Guidance for Latent Diffusion-Based Data Obfuscation

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read CLOAK guides a latent diffusion model to rewrite sensor data so that useful inferences survive while private attributes fade to random-guess levels.

desk verdict Solid, well-engineered obfuscation paper; central claim holds for weakly correlated attributes, but the white-listing guarantee is real only under assumptions the authors themselves show can fail. read the letter →

arxiv 2512.12086 v2 pith:3CWGQIJY submitted 2025-12-12 cs.LG cs.CR

classification cs.LGcs.CR
keywords dataobfuscationlatentdiffusionmodelscontrastivelearningclassifierguidanceattributeinferenceattacksprivacy-utilitytrade-offsensortime-seriesedgeIoTdeployment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes CLOAK, a method that transforms raw sensor time-series or images into obfuscated versions that keep the information needed for a desired inference (e.g., activity recognition) while suppressing information about private attributes (e.g., gender or age). The core idea is to condition a latent diffusion model with a public-attribute representation learned through contrastive sampling, which makes the representation informative about the public attribute yet nearly independent of the private one, and then to subtract the private attribute's influence using a negated classifier during sampling. The authors report that on four time-series datasets and a facial-image dataset, CLOAK consistently beats prior GAN- and diffusion-based obfuscators on the privacy-utility trade-off, improving desired inference accuracy by up to 7.21% and reducing intrusive inference accuracy by up to 5.76% relative to baselines. It also runs in about 203 milliseconds per segment on an edge device, suggesting real-time on-device obfuscation is feasible.

What carries the argument

The score-estimator combination in Equation (12): (1+wU)epsilon_theta(zU) - wU epsilon_theta(empty) + wS sqrt(1-alpha_t) grad_z log p_eta(S|zU, z0_hat). The first two terms implement classifier-free guidance with the contrastively learned public representation zU; the third term implements negated classifier guidance applied to the clean latent prediction z0_hat via universal guidance. The contrastive encoder, trained with InfoNCE where positives share the public label and negatives do not, is what gives the disentanglement property that makes positive and negative conditioning non-interfering.

What would settle it

Construct a synthetic dataset where public and private attributes are perfectly correlated (every public class has exactly one private label). If CLOAK still drives private-attribute inference to random guessing while keeping public-attribute utility high, the implicit-disentanglement claim survives; if utility collapses or privacy loss stays high, the claim fails.

Watch

Extended reading notes

Core claim

CLOAK's central claim is that a single latent diffusion model, trained once, can serve many users with different privacy needs: two guidance terms set at sampling time decide the trade-off. Positive conditioning comes from a contrastive encoder trained so that samples sharing the public attribute map close together regardless of their private attribute; this implicit disentanglement makes the representation a 'white list' that preserves the public attribute and, as a side effect, protects unspecified attributes. Negative conditioning comes from an auxiliary classifier for the private attribute whose gradient is subtracted from the noise prediction, pushing the generated data away from the pr

Load-bearing premise

The disentanglement guarantee rests on the assumption that the training data contains enough samples that share the public attribute but differ in the private one, and that public and private attributes are not strongly correlated; the paper's own face-image results show that when they are strongly correlated, both utility and privacy degrade sharply.

Editorial extensions

If this is right

  • Users can tune two hyperparameters (wU, wS) to navigate the privacy-utility trade-off per user or per session without retraining.
  • Because the contrastive encoder never sees private labels, CLOAK incidentally suppresses unspecified attributes, offering a white-listing form of protection.
  • Multiple private attributes can be protected at once by summing multiple negated guidance gradients, with only a modest utility cost.
  • Operating in a latent space shrinks model size and sampling time enough for near-real-time obfuscation on resource-constrained IoT devices.
  • The same recipe works across motion sensors, WiFi channel state information, audio embeddings, and facial images, suggesting a general obfuscation template.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The contrastive sampling strategy could serve as a general attribute-masking mechanism for latent diffusion, useful beyond privacy for tasks like debiasing generative models.
  • The paper's evidence that strongly entangled attributes (user ID vs. age) degrade the trade-off points to a practical diagnostic: estimate mutual information between public and private attributes before deployment, because high entanglement is exactly where the guarantee weakens.
  • A natural extension would be a verifiable privacy certificate, such as a neural estimator that bounds I(S; X') below a threshold, rather than relying solely on empirical attack accuracy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes CLOAK, a latent diffusion-based data obfuscation framework for protecting attribute privacy in sensor time-series and images. The method trains a contrastive encoder to extract a public-attribute representation z_U, conditions a latent diffusion model on z_U via classifier-free guidance (CCFG), and adds negated classifier guidance to suppress a user-specified private attribute S. The authors claim that CCFG implicitly disentangles z_U from S, providing white-listing protection for unspecified attributes, and that the resulting system achieves state-of-the-art privacy-utility trade-offs on five datasets while being lightweight enough for edge deployment. Experiments compare against ObscureNet, MaSS, and PrivDiffuser, and include a Jetson deployment latency measurement.

Significance. If the claims hold, CLOAK is a practically useful contribution: it is the first latent-diffusion obfuscation framework of its kind for sensor data, with flexible per-user privacy-utility control at sampling time and a reported 4.72x speedup over PrivDiffuser. The empirical work is extensive, covers multiple modalities, uses released code for baselines where possible, and includes statistical significance tests. The disentanglement claim is supported by MINE estimates as an internal check. However, the central privacy-utility claim is conditional on a weak public-private correlation assumption that is violated by the paper's own Adience age-obfuscation experiment, so the broad conclusion 'consistently achieves state-of-the-art' is not fully supported.

major comments (3)
  1. [Eqs. (8)-(10), (12)] The disentanglement argument in §IV-B1 targets I(z_U;S|U), but the evaluated privacy loss L_S is marginal inference accuracy. When U and S are correlated, reducing I(z_U;S|U) does not reduce I(z_U;S), because an adversary can recover S through U. This is precisely what the Adience results show: MI(ID,age)=0.872 vs MI(ID,gender)=0.097, and age obfuscation yields 84.22% utility with 20.40% privacy loss, versus 96.61% utility with 0.09% loss for gender. The unspecified age attribute after gender obfuscation still has a 42.14% privacy loss. The text in §V-G acknowledges the effect but the abstract and conclusion nonetheless state that CLOAK 'consistently achieves state-of-the-art privacy-utility trade-offs' and the white-listing advantage is presented as general. Please narrow these claims to settings with weak public-private correlation, or add a conditional-on-U privacy metric/attack that
  2. [Eqs. (8)-(10), (12)] The negated classifier guidance derivation is internally inconsistent and mathematically unsound as written. Eq. (9) gives a negative sign before ∇ log p_η(S|U,z_t), but Eq. (10) and Eq. (12) have +w_S√(1-ᾱ_t)∇ log p_η(S|z_U,z_t). Since increasing w_S is reported to suppress the private attribute (§V-H, Fig. 4), the sign in Eq. (10) contradicts both Eq. (9) and the described behavior. Moreover, Eq. (8) sets p(¬S|U,z) ∝ 1/p(S|U,z), which is not a proper probabilistic negation; for a classifier p(S|U,z), the correct score for p(¬S|U,z) is -[p/(1-p)]∇ log p(S|U,z), not -∇ log p(S|U,z). Please correct the equations, state the approximation used (e.g., low-probability regime), or redesign the auxiliary model to directly provide a negated-condition score.
  3. [§V-A5, §V-G, Table III] The paper's own results define the scope of the method. On Adience, when the private attribute is strongly entangled with the public attribute (user ID vs age), utility drops from ~99% raw to 84.22% and privacy loss is 20.40%, which is far from the near-random-guessing results on other datasets. This is a load-bearing boundary condition: the method's advantage is specifically demonstrated for weakly entangled attribute pairs. The manuscript needs to explicitly state this limitation in the abstract and conclusion, and ideally provide a pre-deployment check (e.g., estimate MI between public and private attributes) as practical guidance. Without this, the stated claims overstate the method's generality.
minor comments (5)
  1. [Abstract/Conclusion] The abstract says 'reducing utility loss by up to 7.21%' while the conclusion says 'improving desired inference accuracy ... by up to 7.21%'. These are equivalent, but the wording is confusing because 'utility loss' elsewhere refers to accuracy degradation. Please align the phrasing.
  2. [§V-F] The privacy loss metric is accuracy deviation from random guessing, which is misleading when the private attribute has imbalanced classes. The paper notes this for AudioMNIST accent (F1: 5.82%) but still presents the accuracy-based privacy loss prominently. Consider reporting balanced accuracy or macro-F1 for multi-class private attributes in the main tables.
  3. [§V-G] MaSS* results are taken from the original paper's reported values rather than reproduced by the authors, as the implementation is unavailable. This is acknowledged, but the comparison should also state that the evaluation models and data splits may differ, which weakens the direct numerical comparison.
  4. [§IV-A] The paper omits how the unconditional noise predictor ϵ_θ(z_t,t) is trained. Classifier-free guidance requires randomly dropping the condition z_U during training; this should be stated explicitly in the methodology.
  5. [§V-D] The grid-search ranges for w_U and w_S are only given for MotionSense (§V-H). For other datasets, only the selected values are reported. A short sensitivity table or text stating that the selected values were not sensitive in a small neighborhood would strengthen the reproducibility and robustness claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported results are measured, guidance equations are standard, and the entanglement caveat is a scope limitation rather than a circular step.

full rationale

I walked the claimed derivation chain in CLOAK. The guidance formulas in Eqs. (7)-(13) are the standard classifier-free and classifier-guidance algebra (following Ho & Salimans, Dhariwal & Nichol, Dong et al., Liu et al., and Bansal et al.); they are not derived from the paper's own results. The privacy and utility numbers in Tables II-III are measured on held-out test data after grid-searching the hyperparameters w_U and w_S in Section V-D; these are tuning weights, not fitted parameters that are later renamed as predictions. The disentanglement and white-listing argument in Section IV-B1 is explicitly framed as intuition resting on two stated assumptions (diverse positive pairs and weak public-private correlation), and Section V-G/Table III actually demonstrates the failure mode when those assumptions break (estimated MI 0.872 for ID-age vs. 0.097 for ID-gender, with age obfuscation dropping to 84.22% utility and 20.40% privacy loss). That is a correctness/scope limitation, not a circular reduction. The only self-citation, [5] (PrivDiffuser), is used as a prior-work baseline and for a comparison point; it is not load-bearing because CLOAK's deployment claim is directly measured on a Jetson (203.5 ms/segment), and the baseline comparison uses code released by the authors. No equation or fitted quantity reduces by construction to the paper's own inputs.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central contribution is an engineering/methodological combination of existing generative-model machinery; the main hand-set knobs are the two guidance weights (grid-searched per dataset) plus a few VAE/diffusion hyperparameters. The load-bearing domain assumption is the weak-correlation/diverse-data condition under which contrastive sampling is argued to disentangle public from private information. No new physical or mathematical entities are introduced.

free parameters (5)
  • w_U (public guidance weight) = 4.5 (MotionSense); 6 (WiFi-HAR, Adience); 2 (AudioMNIST); 4 (MobiAct)
    Grid-searched per dataset to maximize the reported utility/privacy trade-off; no validation split is described, so the test set may have influenced the choice.
  • w_S (negated private guidance weight) = 0.008 (MotionSense); 0.005 (WiFi-HAR); 0.07 (AudioMNIST); 0.05 (MobiAct); 0.5 (Adience)
    Grid-searched per dataset to control private-attribute suppression; a key knob in the claimed trade-off flexibility.
  • VAE KL divergence weight = 1e-7 (AudioMNIST); 1e-6 (all others)
    Chosen by hand per dataset; affects the latent space used by the diffusion model.
  • Latent dimension = 60 (all datasets)
    Fixed architecture choice for the VAE latent space; not swept, but determines capacity and efficiency.
  • DDIM sampling steps = 50 (all datasets)
    Chosen to balance sampling speed and quality; the efficiency comparison with PrivDiffuser uses this setting.
assumptions (6)
  • standard math DDPM forward process and score-matching objective (Eqs. 1–3) and the LDM extension (Eqs. 4–5) from [28,21].
    Background used as-is; not derived in this paper. The latent diffusion formulation is the substrate for CLOAK's efficiency claim.
  • standard math Classifier-free guidance combination (Eq. 7) from [30].
    Used directly for public-attribute conditioning; assumed to steer sampling toward the public attribute.
  • standard math Compositional/negated classifier guidance (Eqs. 8–10) from [31,32] and universal guidance via \hat z0 (Eqs. 11–12) from [33].
    Used to combine conditions and to evaluate the auxiliary privacy classifier on clean predictions; standard background.
  • domain assumption Threat model: honest-but-curious adversary without metadata and unaware whether data is obfuscated; public attribute U time-invariant (Section III).
    Defines the security goal. If the adversary knows obfuscation is applied and has side information, privacy guarantees may weaken.
  • domain assumption Diverse training data and weak public-private correlation; positive pairs share U but differ in S, so InfoNCE is argued to minimize I(z_U;S|U) (Section IV-B1).
    This is the load-bearing premise for disentanglement. The paper acknowledges it as practical assumption, and its own Adience experiments show failure when correlation is strong.
  • domain assumption The auxiliary privacy model eta trained on clean latent representations remains valid for noisy latents when applied to \hat z0 (universal guidance).
    Universal guidance is adopted to bridge the train/test distribution gap; if the approximation is poor, negated guidance could be misdirected.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLOAK: Contrastive Guidance for Latent Diffusion-Based Data Obfuscation." pith.science (2026). https://pith.science/paper/3CWGQIJY

@misc{pith2026251212086,
  author       = {Pith},
  title        = {Pith review of: CLOAK: Contrastive Guidance for Latent Diffusion-Based Data Obfuscation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CWGQIJY}},
  note         = {Machine review of arXiv:2512.12086}
}
read the original abstract

Data obfuscation is a promising technique for mitigating attribute inference attacks by semi-trusted parties with access to time-series data emitted by sensors. Recent advances leverage conditional generative models together with adversarial training or mutual information-based regularization to balance data privacy and utility. However, these methods often require modifying the downstream task, struggle to achieve a satisfactory privacy-utility trade-off, or are computationally intensive, making them impractical for deployment on resource-constrained mobile IoT devices. We propose Cloak, a novel data obfuscation framework based on latent diffusion models. In contrast to prior work, we employ contrastive learning to extract disentangled representations, which guide the latent diffusion process to retain useful information while concealing private information. This approach enables users with diverse privacy needs to navigate the privacy-utility trade-off with minimal retraining. Extensive experiments on four public time-series datasets, spanning multiple sensing modalities, and a dataset of facial images demonstrate that Cloak consistently outperforms state-of-the-art obfuscation techniques, reducing utility loss by up to 7.21% and privacy loss by up to 5.76%, and is well-suited for deployment in resource-constrained settings.

Figures

Figures reproduced from arXiv: 2512.12086 by the authors.

Figure 1
Figure 1. Overview of CLOAK’s architecture. suited for accommodating diverse privacy needs or adapting to individual user preferences. Recent work by Yang et al. [5] introduced PrivDiffuser, a diffusion model-based obfuscation framework that achieves a strong privacy-utility trade-off, outperforming GAN-based models. However, its obfuscation process is computationally intensive, making it less suitable for deployment on resou… view at source ↗
Figure 2
Figure 2. Illustration of gender obfuscation on Adience. User ID is the public attribute, age is an unspecified attribute. First row: original images, second row: [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Illustration of age obfuscation on Adience. User ID is the public attribute, gender is an unspecified attribute. First row: original images, second row: [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: CLOAK’s privacy-utility trade-off on MotionSense disentanglement [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Estimated mutual information between contrastively learned [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 4 linked inside Pith

  1. [1]

    Mobile sensor data anonymization,

    M. Malekzadehet al., “Mobile sensor data anonymization,” inProceed- ings of the International Conference on Internet-of-Things Design and Implementation, 2019, pp. 49–58

  2. [2]

    Dp-vae: Human-readable text anonymization for online re- views with differentially private variational autoencoders,

    B. Weggenmann, V . Rublack, M. Andrejczuk, J. Mattern, and F. Ker- schbaum, “Dp-vae: Human-readable text anonymization for online re- views with differentially private variational autoencoders,” inProceed- ings of the ACM Web Conference 2022, 2022, pp. 721–731

  3. [3]

    Blind-touch: Homomorphic encryption-based distributed neural network inference for privacy- preserving fingerprint authentication,

    H. Choi, S. S. Woo, and H. Kim, “Blind-touch: Homomorphic encryption-based distributed neural network inference for privacy- preserving fingerprint authentication,” inProceedings of the AAAI con- ference on artificial intelligence, vol. 38, no. 20, 2024, pp. 21 976– 21 985

  4. [4]

    ObscureNet: Learning attribute-invariant latent representation for anonymizing sensor data,

    O. Hajihassnaiet al., “ObscureNet: Learning attribute-invariant latent representation for anonymizing sensor data,” inProceedings of the Inter- national Conference on Internet-of-Things Design and Implementation, 2021, pp. 40–52

  5. [5]

    Privdiffuser: Privacy-guided diffusion model for data obfuscation in sensor networks,

    X. Yang and O. Ardakanian, “Privdiffuser: Privacy-guided diffusion model for data obfuscation in sensor networks,”Proceedings on Privacy Enhancing Technologies, vol. 2025, no. 4, pp. 40–55, 2025

  6. [6]

    Privacy adversarial network: representation learning for mobile data privacy,

    S. Liuet al., “Privacy adversarial network: representation learning for mobile data privacy,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 3, no. 4, pp. 1–18, 2019

  7. [7]

    DeepObfuscator: Obfuscating intermediate representations with privacy-preserving adversarial learning on smartphones,

    A. Liet al., “DeepObfuscator: Obfuscating intermediate representations with privacy-preserving adversarial learning on smartphones,” inPro- ceedings of the International Conference on Internet-of-Things Design and Implementation, 2021, pp. 28–39

  8. [8]

    TIPRDC: Task-independent privacy-respecting data crowd- sourcing framework for deep learning with anonymized intermediate representations,

    A. Liet al., “TIPRDC: Task-independent privacy-respecting data crowd- sourcing framework for deep learning with anonymized intermediate representations,” inProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020, pp. 824– 832

Show all 41 references
  1. [9]

    Olympus: Sensor privacy through utility aware ob- fuscation

    N. Ravalet al., “Olympus: Sensor privacy through utility aware ob- fuscation.”Proceedings on Privacy Enhancing Technologies, vol. 2019, no. 1, pp. 5–25, 2019

  2. [10]

    MaSS: Multi-attribute selective suppression for utility- preserving data transformation from an information-theoretic perspec- tive,

    Y . Chenet al., “MaSS: Multi-attribute selective suppression for utility- preserving data transformation from an information-theoretic perspec- tive,” inProceedings of the 41st International Conference on Machine Learning, 2024, pp. 6519–6538

  3. [11]

    Adversarially learned representations for information obfuscation and inference,

    M. Bertran, N. Martinez, A. Papadaki, Q. Qiu, M. Rodrigues, G. Reeves, and G. Sapiro, “Adversarially learned representations for information obfuscation and inference,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 614–623

  4. [12]

    Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset,

    Z. Wu, H. Wang, Z. Wang, H. Jin, and Z. Wang, “Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 4, pp. 2126–2139, 2020

  5. [13]

    Spact: Self-supervised privacy preservation for action recognition,

    I. R. Dave, C. Chen, and M. Shah, “Spact: Self-supervised privacy preservation for action recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 20 164–20 173

  6. [14]

    Disentangled and controllable face image generation via 3d imitative-contrastive learning,

    Y . Denget al., “Disentangled and controllable face image generation via 3d imitative-contrastive learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 5154– 5163

  7. [15]

    Learning disentangled representation by exploiting pretrained generative models: A contrastive learning view,

    X. Ren, T. Yang, Y . Wang, and W. Zeng, “Learning disentangled representation by exploiting pretrained generative models: A contrastive learning view,” inInternational Conference on Learning Repre- sentations, 2022. [Online]. Available: https://openreview.net/pdf?id=j- 63FSNcO5a

  8. [16]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

  9. [17]

    Uncovering the disentanglement capability in text-to- image diffusion models,

    Q. Wu, Y . Liu, H. Zhao, A. Kale, T. Bui, T. Yu, Z. Lin, Y . Zhang, and S. Chang, “Uncovering the disentanglement capability in text-to- image diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 1900–1910

  10. [18]

    Disdiff: Unsupervised disentanglement of diffusion probabilistic models,

    T. Yang, Y . Wang, Y . Lu, and N. Zheng, “Disdiff: Unsupervised disentanglement of diffusion probabilistic models,”Advances in Neural Information Processing Systems, vol. 36, pp. 69 130–69 156, 2023

  11. [19]

    Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models,

    Y . Dalva and P. Yanardag, “Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 24 209–24 218

  12. [20]

    Deadiff: An efficient stylization diffusion model with dis- entangled representations,

    T. Qiet al., “Deadiff: An efficient stylization diffusion model with dis- entangled representations,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 8693–8702

  13. [21]

    High-resolution image synthesis with latent diffu- sion models,

    R. Rombachet al., “High-resolution image synthesis with latent diffu- sion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

  14. [22]

    Learning transferable visual models from natural lan- guage supervision,

    A. Radfordet al., “Learning transferable visual models from natural lan- guage supervision,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 8748–8763

  15. [23]

    Diffusegae: controllable and high-fidelity image manipulation from disentangled representation,

    Y . Leng, Q. Huang, Z. Wang, Y . Liu, and H. Zhang, “Diffusegae: controllable and high-fidelity image manipulation from disentangled representation,” inProceedings of the 5th ACM International Conference on Multimedia in Asia, 2023, pp. 1–7

  16. [24]

    Diffusion model with cross attention as an inductive bias for disentanglement,

    T. Yang, C. Lan, Y . Lu, and N. Zheng, “Diffusion model with cross attention as an inductive bias for disentanglement,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  17. [25]

    Factorized diffusion autoencoder for unsuper- vised disentangled representation learning,

    A. Wu and W.-S. Zheng, “Factorized diffusion autoencoder for unsuper- vised disentangled representation learning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 6, 2024, pp. 5930– 5939

  18. [26]

    Re-imagine the negative prompt algorithm: Transform 2D diffusion into 3D, alleviate Janus problem and beyond,

    M. Armandpouret al., “Re-imagine the negative prompt algorithm: Transform 2D diffusion into 3D, alleviate Janus problem and beyond,” arXiv preprint arXiv:2304.04968, 2023

  19. [27]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,”Advances in Neural Information Processing Systems, vol. 34, pp. 8780–8794, 2021

  20. [28]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in Neural Information Processing Systems, vol. 33, pp. 6840– 6851, 2020

  21. [29]

    Auto-encoding variational bayes,

    D. P. Kingma, M. Wellinget al., “Auto-encoding variational bayes,” 2013

  22. [30]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

  23. [31]

    Towards test-time refusals via concept negation,

    P. Donget al., “Towards test-time refusals via concept negation,” Advances in Neural Information Processing Systems, vol. 36, 2024

  24. [32]

    Compositional visual generation with composable diffu- sion models,

    N. Liuet al., “Compositional visual generation with composable diffu- sion models,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 423–439

  25. [33]

    Universal guidance for diffusion models,

    A. Bansalet al., “Universal guidance for diffusion models,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 843–852

  26. [34]

    Human daily activity and fall recognition using a smartphone’s acceleration sensor,

    C. Chatzakiet al., “Human daily activity and fall recognition using a smartphone’s acceleration sensor,” inInternational Conference on Information and Communication Technologies for Ageing Well and e- Health. Springer, 2016, pp. 100–118

  27. [35]

    A dataset for wi-fi-based human activity recognition in line-of-sight and non-line-of-sight indoor environments,

    A. Baha’Aet al., “A dataset for wi-fi-based human activity recognition in line-of-sight and non-line-of-sight indoor environments,”Data in Brief, vol. 33, p. 106534, 2020

  28. [36]

    Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark,

    S. Becker, J. Vielhaben, M. Ackermann, K.-R. M ¨uller, S. Lapuschkin, and W. Samek, “Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark,”Journal of the Franklin Institute, vol. 361, no. 1, pp. 418–428, 2024

  29. [37]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

  30. [38]

    Age and gender estimation of unfiltered faces,

    E. Eidinger, R. Enbar, and T. Hassner, “Age and gender estimation of unfiltered faces,”IEEE Transactions on information forensics and security, vol. 9, no. 12, pp. 2170–2179, 2014

  31. [39]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inProceedings of the 18th Interna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2015, pp. 234–241

  32. [40]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502, 2020

  33. [41]

    Mutual information neural estimation,

    M. I. Belghaziet al., “Mutual information neural estimation,” inInter- national conference on machine learning. PMLR, 2018, pp. 531–540

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.