Pith. sign in

REVIEW 3 major objections 4 minor 49 references

FedTaste: Topology-Aware Structural Transfer for Multimodal Federated Learning with Missing Modalities

T0 review · 3 major / 4 minor · reviewed 2026-07-31 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FedTaste transfers group-level semantic structure from full-modality teachers to missing-modality students, avoiding imputation and external data.

desk verdict A plausible topology-transfer framework for missing-modality MFL, but the main experiments are invalidated by an unfair CLIP-based comparison and an undefined retrieval objective. read the letter →

arxiv 2607.23245 v1 pith:Y6C5AMNX submitted 2026-07-25 cs.MM cs.LG

Haochen Liang , Jie Zhang , Hideya Ochiai This is my paper
classification cs.MMcs.LG
keywords multimodalfederatedlearningmissingmodalitiesstructuraltransfersemantictopologyrepresentationdriftModality-AdaptivePromptsspectralconsistencynon-IID
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 claims that under missing modalities and non-IID data, aligning second-order group-level semantic relations is more robust than aligning first-order feature coordinates. FedTaste has full-modality teachers construct local affinity graphs over semantic groups, the server consolidates these into a confidence-aware global blueprint, and missing-modality students adapt lightweight prompts to match that blueprint plus the graph's spectrum. If this works as claimed, missing-modality clients can join multimodal federated learning without generative imputation or public auxiliary data, and with lower communication cost.

What carries the argument

The engine is the relational graph over semantic groups: teacher clients build affinity matrices from fused image-text prototypes, upload them with Laplacian eigenvalues; the server filters edges by support frequency and cross-client variance to form a masked blueprint; students minimize a contrastive task loss plus a topology transfer loss on masked edges and a spectral loss on eigenvalues, using only learnable prompts on frozen backbones.

What would settle it

Take an image-only client from Flickr30k where each image is its own group — the positive set P(n) is empty for most samples — and run the reported protocol; if the topology loss or task loss cannot be computed (or R@1 sum collapses), the method's claimed retrieval results are not reproducible as described.

Watch

Extended reading notes

Core claim

The central claim is that stable group-level semantic topology can be distilled from full-modality clients and used to align partial representations of missing-modality clients, so that absolute coordinate drift under Non-IID conditions no longer breaks cross-modal alignment. The paper argues that pairwise affinities among group prototypes and the low-frequency spectrum of the normalized Laplacian survive representation drift, and that a confidence-filtered global blueprint of these quantities provides a reliable target for student-side adaptation.

Load-bearing premise

Every missing-modality client must have group labels for its single modality and enough same-group samples for the contrastive positive set; for retrieval benchmarks where images are unique instances, the paper specifies no way to form such groups, leaving Eq. (10) undefined.

Editorial extensions

If this is right

  • Missing-modality clients can be added to a multimodal federation without generative models or public data, lowering privacy risk and compute.
  • Communication drops to 174.58 MB versus 371.26 MB for the strongest first-order alignment baseline on Flickr30k.
  • The method degrades gracefully when full-modality teachers are scarce, keeping a clear lead even when image-text pairs are limited.
  • Spectral matching stabilizes macro-cluster structure, which could extend alignment to other modality combinations beyond image and text.

Reading between the lines

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

  • If second-order structure is as drift-resistant as claimed, the same pipeline could align new modalities (audio, sensors) without any paired multimodal data on the student side.
  • The group-label requirement is a hidden constraint; for instance-based retrieval data where each image is unique, the paper does not specify how a group or positive set forms, so the cleanest test bed is class-labeled data.
  • One could directly test drift-resistance by measuring how much pairwise affinity matrices change under simulated local shifts relative to raw coordinates.
  • The confidence-aware variance filter suggests a communication schedule where only cross-client stable edges need to be re-transmitted each round.
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 / 4 minor

Summary. FedTaste proposes a topology-aware method for multimodal federated learning when clients lack image or text modality. Full-modality teachers keep a frozen CLIP backbone, compute group-level prototypes and pairwise affinity matrices, and upload these together with normalized-Laplacian eigenvalues to the server. The server builds a confidence-filtered global structural blueprint, which missing-modality students use as a target while updating only lightweight prompts; a spectral-consistency loss and a group-supervised contrastive loss complete the student objective. Experiments on Flickr30k/MS-COCO retrieval and on medical classification report consistent improvements over baselines across heterogeneity, participation, and modality-skew settings, together with lower communication cost, ablations, and hyperparameter sensitivity studies.

Significance. If correct, the central idea — transferring stable second-order group-level topology instead of first-order features — would be a valuable alternative to generative imputation and external-data-based alignment in multimodal federated learning. The paper offers a concrete mechanism (structural payloads, confidence masking, prompt-based student adaptation) and includes ablation and communication-cost analyses. However, the empirical claim is not currently supported: the student objective is undefined for the main retrieval experiments as written, and the baseline comparison is confounded by the exclusive use of a frozen CLIP backbone. The significance of the contribution, therefore, cannot be assessed from the present manuscript.

major comments (3)
  1. [§3.5, Eq. (10)] The student task loss is undefined for image-only clients on the retrieval datasets used in Tables 2 and 3. Section 3.1 defines a group as 'an image-centered pairing unit or its associated semantic anchor.' In Flickr30k/MS-COCO, an image-only client has exactly one image per group, so no other sample in its local batch has the same group label; hence P(n) is empty and Eq. (10) divides by |P(n)|. The paper also never states how y_n is obtained for images (which have no class labels) or whether the Dirichlet partition preserves multiple captions of the same image on the same client. This makes the reported retrieval results impossible to reproduce from the described algorithm. It is a load-bearing correctness failure, not a tuning issue.
  2. [§4.1, Tables 2–3] FedTaste is the only method evaluated with frozen CLIP ViT-B/32 backbones while baselines use their original recommended architectures. The large gains in Tables 2 and 3 therefore conflate the proposed topology-transfer mechanism with the advantage of a strong pretrained backbone. A fair comparison would run baselines on the same frozen CLIP backbone or at least report CLIP-initialized baseline variants. This is essential because the paper's central claim is superiority of structural transfer, not CLIP initialization.
  3. [Table 5, §4.4] The 'domain shift' table reproduces exactly the default-setting values from Table 2 for CreamFL (74.83), FedIoT (85.51), FedCola (91.96), and FedTaste (99.64), yet the text describes a medical federation built from OrganCMNIST and Medical Abstracts. As printed, the table cannot support the claimed robustness under domain shift; either the experiment was not conducted as described or the table is mislabeled. The claim in §4.4 therefore needs to be re-evaluated.
minor comments (4)
  1. [§3.5, Eq. (12)] The normalization denominator uses ||...||_0, but the operand is the difference matrix M_i - M_hat, not a binary mask. This is not the count of active edges; use the sum of entries of the binary mask instead.
  2. [§4.5] Hyperparameters K and beta are selected by sensitivity analysis on the same benchmark used for the main results, and no standard deviations or multiple seeds are reported. This limits the strength of the conclusions about robustness and statistical significance.
  3. [§4.1] Implementation details such as MASP insertion layers, baseline configurations, and per-round communication breakdowns are deferred to a supplementary material that is not present in the arXiv version, impeding reproducibility.
  4. [§3.4, Eq. (6)] The per-row retention count K_n is not specified, nor is its relationship to the global confidence mask. Please define how K_n is chosen.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: FedTaste's topology transfer is standard relational distillation from frozen CLIP; the main issues are a non-load-bearing self-citation and a definitional gap in Eq. (10) that is a correctness concern, not circularity.

full rationale

The derivation chain is not circular. Teacher-side joint prototypes (Eqs. 1-2) are built from frozen CLIP features, the affinity graph (Eq. 3) and Laplacian spectrum (Eqs. 4-5) are functions of those prototypes, and the server-side blueprint (Eqs. 6-8) is a confidence-weighted aggregation of teacher affinities and spectra. Student adaptation (Eqs. 10-15) minimizes a local contrastive task loss, a squared-Frobenius topology loss against the blueprint, and a spectral-consistency loss. The supervision signal (the blueprint) is computed from teacher-local data independently of the student's local task labels; the student does not feed back into the blueprint, so this is standard structural distillation rather than a self-justifying prediction. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work. The only self-citation, Ref. [17] (FedCoop by the first author), appears in the related-work critique of generative imputation and is not load-bearing for the central claim. The stability of second-order relations is asserted with external citations [11,36] and functions as an assumption rather than a derived result; an assumption is not circularity. Two caveats should be weighed separately: K and beta are selected by sensitivity on the benchmark itself, which is hyperparameter tuning rather than circularity; and Eq. (10) requires a non-empty positive set P(n), which is undefined for an image-only retrieval client if each image-centered group contains a single image — that is a correctness/reproducibility defect, not a circular reduction. For these reasons, the paper's central derivation is not circular, and the score is low.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claims rest on empirical stability assumptions about second-order structure and on a shared group-labeling scheme that is not fully specified for retrieval. No new physical entities are introduced; MASP and the global blueprint are algorithmic constructs, not independently evidenced entities.

free parameters (8)
  • Topology weight β = 0.5
    Selected by sensitivity analysis; inverted-U trend in Fig. 4(b), peak at 0.5 on the same benchmark as the main results.
  • Spectral retainment K = 5
    Retained eigenvalues of the normalized Laplacian; selected by sensitivity analysis with peak at K=5.
  • Warm-up horizon R_w = 10
    Chosen to schedule the topology weight; no sensitivity study reported.
  • Spectral consistency weight γ = not reported
    Appears in Eq. (15), but no numerical value or tuning procedure is given in the paper.
  • Regularization weight η = not reported
    Appears in Eq. (15) for Frobenius-norm prompt regularization; value not given.
  • Affinity temperature τ_s = not reported
    Used in Eq. (3) for the affinity matrix; value not given.
  • Task temperature τ_t = not reported
    Used in Eq. (10) for the contrastive loss; value not given.
  • Variance threshold ρ and top-K_n edge selection = not reported
    Used in Eq. (6) to filter the global blueprint; neither value is specified.
assumptions (4)
  • domain assumption Second-order relational structure among semantic groups is more stable than first-order feature coordinates under non-IID representation drift.
    Central motivation stated in §1 and §3.2; supported only by citations, not by a theorem or controlled experiment isolating this claim.
  • domain assumption A unified, known set of semantic groups G exists, and each client's single-modality samples carry group labels usable for contrastive learning.
    Required by Eq. (10) and the student-side graph construction; not concretely defined for retrieval datasets where images are unique instances.
  • domain assumption Frozen CLIP encoders provide a compatible shared embedding space for both modalities and for target domains, including medical OOD settings.
    All teacher and student features come from CLIP (ViT-B/32); if CLIP is not appropriate for the target domain, the whole structural blueprint is built on mismatched semantics.
  • domain assumption Batch-level group prototypes are representative enough to build reliable local affinity matrices.
    Eq. (11) averages only batch samples per group; with few samples per group or unique instances, the local graph may be dominated by noise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedTaste: Topology-Aware Structural Transfer for Multimodal Federated Learning with Missing Modalities." pith.science (2026). https://pith.science/paper/Y6C5AMNX

@misc{pith2026260723245,
  author       = {Pith},
  title        = {Pith review of: FedTaste: Topology-Aware Structural Transfer for Multimodal Federated Learning with Missing Modalities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y6C5AMNX}},
  note         = {Machine review of arXiv:2607.23245}
}
read the original abstract

Multimodal Federated Learning is often challenged by arbitrary modality missingness and Non-IID data distributions, which lead to severe representation drift and hinder effective collaboration across clients. Existing methods typically rely on generative imputation, external auxiliary data, or isolated unimodal training to bridge modality gaps, often incurring substantial communication and computational costs as well as potential privacy risks. To address these limitations, we propose FedTaste, a parameter-efficient framework for topology-aware structural transfer in Multimodal Federated Learning with missing modalities. Instead of aligning fragile first-order features, FedTaste focuses on more stable group-level semantic relations. Specifically, FedTaste leverages frozen foundation models to extract a joint multimodal topology from full-modality clients, which is then consolidated by the server into a global structural blueprint. To adapt clients with missing modalities, we introduce Modality-Adaptive Structural Prompts together with spectral consistency regularization, enabling lightweight branch-specific adaptation that aligns local partial representations with the shared blueprint. In this way, FedTaste avoids explicit modality imputation while preserving shared semantic structure across clients. Extensive experiments demonstrate that FedTaste consistently achieves superior performance across multiple datasets and challenging Non-IID settings, while substantially reducing communication overhead compared with existing methods.

Figures

Figures reproduced from arXiv: 2607.23245 by the authors.

Figure 1
Figure 1. Two major challenges in federated multimodal [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Motivation of the FedTaste framework. Existing paradigms for missing-modality MFL suffer from three limitations: [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of the FedTaste framework. FedTaste addresses federated multimodal learning with missing modalities [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Hyperparameter sensitivity analysis of the retained [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Cross-modal affinity heatmaps under first-order [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: t-SNE visualization of latent representations un [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 2 linked inside Pith

  1. [1]

    Ziteng Bao, Manuel Aleixandre, Shoichi Hasegawa, and Takamichi Nakamoto

  2. [2]

    Zhibo Cao, Kuangrong Hao, Lingguang Hao, Bing Wei, and Lihong Ren. 2026. Heterogeneous Multimodal Federated Learning with Missing Modality via Mask- Restoration and Self-Guidance.IEEE Transactions on Multimedia(2026)

    2026

  3. [3]

    Liwei Che, Jiaqi Wang, Yao Zhou, and Fenglong Ma. 2023. Multimodal federated learning: A survey.Sensors23, 15 (2023), 6986

    2023

  4. [4]

    Jiayi Chen and Aidong Zhang. 2024. FedMBridge: Bridgeable multimodal feder- ated learning. InForty-first International Conference on Machine Learning

    2024

  5. [5]

    Sijia Chen and Baochun Li. 2022. Towards optimal multi-modal federated learning on non-IID data with hierarchical gradient blending. InIEEE INFOCOM 2022-IEEE conference on computer communications. IEEE, 1469–1478

    2022

  6. [6]

    Suo Chen, Yang Xu, Hongli Xu, Zhenguo Ma, and Zhiyuan Wang. 2024. Enhancing decentralized and personalized federated learning with topology construction. IEEE Transactions on Mobile Computing23, 10 (2024), 9692–9707

    2024

  7. [7]

    Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Pi- otr Dollár, and C Lawrence Zitnick. 2015. Microsoft coco captions: Data collection and evaluation server.arXiv preprint arXiv:1504.00325(2015)

    arXiv 2015

  8. [8]

    Chenyuan Feng, Daquan Feng, Guanxin Huang, Zuozhu Liu, Zhenzhong Wang, and Xiang-Gen Xia. 2024. Robust privacy-preserving recommendation systems driven by multimodal federated learning.IEEE Transactions on Neural Networks and Learning Systems36, 5 (2024), 8896–8910

    2024

Show all 49 references
  1. [9]

    Tiantian Feng, Digbalay Bose, Tuo Zhang, Rajat Hebbar, Anil Ramakrishna, Rahul Gupta, Mi Zhang, Salman Avestimehr, and Shrikanth Narayanan. 2023. Fedmultimodal: A benchmark for multimodal federated learning. InProceedings of the 29th ACM SIGKDD conference on knowledge discover...

    2023

  2. [10]

    Chih-fan Hsu, Yi-chen Li, Chung-chi Tsai, Jian-kai Wang, and Cheng-hsin Hsu

  3. [11]

    Ke Hu, Liyao Xiang, Peng Tang, and Weidong Qiu. 2026. FedTopo: Topology- Informed Representation Alignment in Federated Learning Under Non-IID Con- ditions. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 21849–21857

    2026

  4. [12]

    Wei Huang, Dexian Wang, Xiaocao Ouyang, Jihong Wan, Jia Liu, and Tianrui Li

  5. [13]

    Meirui Jiang, Zirui Wang, and Qi Dou. 2022. Harmofl: Harmonizing local and global drifts in federated learning on heterogeneous medical images. InProceed- ings of the AAAI conference on artificial intelligence, Vol. 36. 1087–1095

    2022

  6. [14]

    Hanna Kavalionak, Emanuele Carlini, Patrizio Dazzi, Luca Ferrucci, Matteo Mor- dacchini, and Massimo Coppola. 2021. Impact of network topology on the con- vergence of decentralized federated learning systems. In2021 IEEE Symposium on Computers and Communications (ISCC). IEEE, 1–6

    2021

  7. [15]

    Multimodal federated learning: Concept, methods, applications and future directions.Information Fusion112 (2024), 102576

    2024

  8. [16]

    Yijing Li, Xiaofeng Tao, Xuefei Zhang, Junjie Liu, and Jin Xu. 2021. Privacy- preserved federated learning for autonomous driving.IEEE Transactions on Intelligent Transportation Systems23, 7 (2021), 8423–8434

    2021

  9. [17]

    Haochen Liang, Chaomeng Chen, Jing Yang, Hui Ma, Junzhe Cao, and Zitong Yu. 2026. FedCoop: Co-optimized Latent Diffusion for Asymmetric Multimodal Federated Learning.Authorea Preprints(2026)

    2026

  10. [18]

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. 2020. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems2 (2020), 429–450

    2020

  11. [19]

    Yi Liu, Cong Wang, and Xingliang Yuan. 2025. Fedmobile: Enabling knowledge contribution-aware multi-modal federated learning with incomplete modalities. InProceedings of the ACM on Web Conference 2025. 2775–2786

    2025

  12. [20]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. InArtificial intelligence and statistics. Pmlr, 1273–1282

    2017

  13. [21]

    Yi-Ming Lin, Yuan Gao, Mao-Guo Gong, Si-Jia Zhang, Yuan-Qiao Zhang, and Zhi-Yuan Li. 2023. Federated learning on multimodal data: A comprehensive survey.Machine Intelligence Research20, 4 (2023), 539–553

    2023

  14. [22]

    Ba Hung Ngo, Doanh C Bui, Nhat-Tuong Do-Tran, and Tae Jong Choi. 2025. Higda: Hierarchical graph of nodes to learn local-to-global topology for semi- supervised domain adaptation. InProceedings of the AAAI conference on artificial intelligence, Vol. 39. 6191–6199

    2025

  15. [23]

    Dinh C Nguyen, Quoc-Viet Pham, Pubudu N Pathirana, Ming Ding, Aruna Senevi- ratne, Zihuai Lin, Octavia Dobre, and Won-Joo Hwang. 2022. Federated learning for smart healthcare: A survey.ACM Computing Surveys (Csur)55, 3 (2022), 1–37

    2022

  16. [24]

    Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, Zhengming Ding, and Chen Chen. 2022. Local learning matters: Rethinking data heterogeneity in federated learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8397–8406

    2022

  17. [25]

    Pranav Poudel, Aavash Chhetri, Prashnna Gyawali, Georgios Leontidis, and Binod Bhattarai. 2025. Multimodal federated learning with missing modalities through feature imputation network. InAnnual Conference on Medical Image Understanding and Analysis. Springer, 289–299

    2025

  18. [26]

    Pranav Poudel, Prashant Shrestha, Sanskar Amgain, Yash Raj Shrestha, Prashnna Gyawali, and Binod Bhattarai. 2024. Car-mfl: Cross-modal augmentation by retrieval for multimodal federated learning with missing modalities. InInterna- tional Conference on Medical Image Computing a...

    2024

  19. [27]

    Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hock- enmaier, and Svetlana Lazebnik. 2015. Flickr30k entities: Collecting region-to- phrase correspondences for richer image-to-sentence models. InProceedings of the IEEE international conference on compute...

    2015

  20. [28]

    Tim Schopf, Daniel Braun, and Florian Matthes. 2022. Evaluating unsupervised text classification: zero-shot and similarity-based approaches. InProceedings of the 2022 6th International Conference on Natural Language Processing and Information Retrieval. 6–15

    2022

  21. [29]

    Jiangming Shi, Shanshan Zheng, Xiangbo Yin, Yang Lu, Yuan Xie, and Yanyun Qu

  22. [30]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  23. [31]

    Guangyu Sun, Matias Mendieta, Aritra Dutta, Xin Li, and Chen Chen. 2024. Towards multi-modal transformers in federated learning. InEuropean Conference on Computer Vision. Springer, 229–246

    2024

  24. [32]

    Min Tan, Yinfu Feng, Lingqiang Chu, Jingcheng Shi, Rong Xiao, Haihong Tang, and Jun Yu. 2023. Fedsea: Federated learning via selective feature alignment for non-iid multimodal data.IEEE Transactions on Multimedia26 (2023), 5807–5822

    2023

  25. [33]

    In Proceedings of the AAAI Conference on Artificial Intelligence, Vol

    Clip-guided federated learning on heterogeneity and long-tailed data. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 14955–14963

  26. [34]

    Haoyue Song, Jiacheng Wang, Jianjun Zhou, and Liansheng Wang. 2024. Tackling modality-heterogeneous client drift holistically for heterogeneous multimodal federated learning.IEEE Transactions on Medical Imaging44, 4 (2024), 1931–1941

    2024

  27. [35]

    Xinliang Wei, Kejiang Ye, Xinghua Shi, Cheng-Zhong Xu, and Yu Wang. 2024. Joint participant and learning topology selection for federated learning in edge clouds.IEEE Transactions on Parallel and Distributed Systems35, 8 (2024), 1456– 1468

    2024

  28. [36]

    Jiajun Wu, Fan Dong, Henry Leung, Zhuangdi Zhu, Jiayu Zhou, and Steve Drew

  29. [37]

    Min Tan, Junchao Ma, Yinfu Feng, Jiajun Ding, Wenwen Pan, Tingting Han, Qian Zheng, Zhenzhong Kuang, and Zhou Yu. 2026. FedAFD: Multimodal Federated Learning via Adversarial Fusion and Distillation.arXiv preprint arXiv:2603.04890 (2026)

    2026

  30. [38]

    Shu Wang, Zhe Qu, Yuan Liu, Shichao Kan, Yixiong Liang, and Jianxin Wang. 2024. Fedmmr: Multi-modal federated learning via missing modality reconstruction. In 2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 1–6

    2024

  31. [39]

    Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. 2023. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification.Scientific data10, 1 (2023), 41. Conference acronym ’XX, June 03–05, 2...

    2023

  32. [40]

    Qiying Yu, Yang Liu, Yimu Wang, Ke Xu, and Jingjing Liu. 2023. Multimodal federated learning via contrastive representation ensemble.arXiv preprint arXiv:2302.08888(2023)

    2023 arXiv

  33. [41]

    Surveys56, 10 (2024), 1–41

    Topology-aware federated learning in edge computing: A comprehensive survey.Comput. Surveys56, 10 (2024), 1–41

    2024

  34. [42]

    Yunlu Yan, Chun-Mei Feng, Yuexiang Li, Ping Li, Rick Siow Mong Goh, Baiying Lei, Weiming Wang, David Dagan Feng, and Lei Zhu. 2025. Federated pseudo modality generation for incomplete multi-modal mri reconstruction.IEEE Journal of Biomedical and Health Informatics(2025)

    2025

  35. [43]

    Chenxiao Yang, Qitian Wu, and Junchi Yan. 2022. Geometric knowledge distil- lation: Topology compression for graph neural networks.Advances in Neural Information Processing Systems35 (2022), 29761–29775

    2022

  36. [44]

    Linlin Zong, Qiujie Xie, Jiahui Zhou, Peiran Wu, Xianchao Zhang, and Bo Xu. 2021. FedCMR: Federated cross-modal retrieval. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1672–1676

    2021

  37. [46]

    Songcan Yu, Kaiming Zhu, Feiyuan Liang, Junbo Wang, Krishna Kant, and Ling Yin. 2026. Robust multimodal federated learning for non-IID multimodal data with incompleteness.Future Generation Computer Systems174 (2026), 107948

    2026

  38. [47]

    Rongyu Zhang, Xiaowei Chi, Wenyi Zhang, Guiliang Liu, Dan Wang, and Fangxin Wang. 2025. Unimodal training-multimodal prediction: Cross-modal federated learning with hierarchical aggregation.IEEE Transactions on Mobile Computing (2025)

    2025

  39. [48]

    Yuchen Zhao, Payam Barnaghi, and Hamed Haddadi. 2022. Multimodal federated learning on iot data. In2022 IEEE/ACM seventh international conference on internet- of-things design and implementation (ioTDI). IEEE, 43–54

    2022

  40. [2024]

    Federated learning using multi-modal sensors with heterogeneous privacy sensitivity levels.ACM Transactions on Multimedia Computing, Communications and Applications20, 11 (2024), 1–27

    2024

  41. [2025]

    IEEE Sensors Journal(2025)

    A Robust Odor Mixture Quantification Method Based on Active Sensing Using Both QCM Frequency Shifts and Resistance Changes of Multiple Harmonics. IEEE Sensors Journal(2025)

    2025

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.