REVIEW 3 major objections 5 minor 48 references
Federated Deep Subspace Clustering
T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read The paper claims that FDSC, a federated deep subspace clustering method that shares only the encoder and aligns the self-expressive matrix with a local k-NN adjacency matrix, outperforms centralized subspace clustering baselines on MNIST…
desk verdict Genuinely new combination of federated learning and deep subspace clustering, but the objective in Eq. (8) is undefined—a load-bearing flaw that invalidates the reported results as written. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the self-expressive matrix $R_i$, an $n\times n$ weight matrix in a fully connected layer of client $i$'s network; ideally its nonzero entries connect each latent point only to other points in the same subspace, giving it block-diagonal structure. The paper's mechanism has three parts: a shared convolutional encoder whose weights are averaged on the server (Eq. (6)), a private self-expressive layer producing $R_i$, and a private decoder for reconstruction. The novelty is the third term in Eq. (8), which pushes $R_i$ toward the local k-NN adjacency matrix $A_i$ so that local neighborhood relations are preserved while the encoder is being federated. Spectral clustering is then applied to the affinity matrix derived from $R_i$ on each client.
What would settle it
Evaluate Eq. (8) directly: attempt to compute the gradient of $\lambda_3(\alpha A_i - \beta R_i)$ with respect to $R_i$; since the term is an $n\times n$ matrix added to scalars, automatic differentiation will fail unless a norm, trace, or elementwise reduction is silently applied. Alternatively, set $\lambda_3 = 0$ and compare FDSC with a centralized graph-regularized deep subspace clustering baseline that uses the same adjacency alignment; if the two perform identically, the federated encoder contributes nothing beyond regularization.
Extended reading notes
Core claim
The central claim is that deep subspace clustering can be federated without losing, and in fact improving, clustering accuracy. Each client keeps its self-expressive layer and decoder private, communicates only the encoder weights to the server, and the server averages the encoders as in FedAvg. The paper further claims that adding a regularization term that aligns the learned self-expressive matrix $R_i$ with a k-NN adjacency matrix $A_i$ built from the client's raw data strengthens the block-diagonal structure of $R_i$, so that points from the same subspace represent each other more cleanly. The experiments in Table 2 are offered as evidence: FDSC2 reaches the best ACC/NMI/AMI/ARI on all four datasets, and the 2D representation and self-expression matrix visualizations show tighter clusters and clearer block diagonals than the centralized DSCN baseline.
Load-bearing premise
The training objective in Eq. (8) is assumed to be a valid scalar loss, but as written it adds the matrix $\lambda_3(\alpha A_i - \beta R_i)$ to scalar reconstruction, norm, and self-expressiveness terms, so the objective has no defined numerical value.
Editorial extensions
If this is right
- If FDSC's claims hold, a client can obtain better local clustering by participating in federation than by running a centralized deep subspace clustering model alone on its own data.
- The adjacency-alignment term is doing real work: FDSC2 (with $\lambda_3 = 10^6$) beats FDSC1 (with $\lambda_3 = 0$) on all four datasets, so preserving local neighborhood structure is what the paper credits for the gain.
- Clustering accuracy improves as the number of clients increases (Table 3) and as batch size increases (Table 4), suggesting that broader participation and more stable local gradients help the federated representation.
- On ORL, COIL20, and COIL100, lower data heterogeneity across clients (larger $q$) yields better clustering, while MNIST behaves oppositely, so the effect of heterogeneity is dataset-dependent.
Reading between the lines
- The comparison with centralized baselines frames FDSC as a privacy-preserving alternative, but the paper does not quantify privacy leakage through the shared encoder; an evident extension is an attack or inversion experiment on the communicated weights.
- The adjacency-alignment idea could transfer to other self-expressive formulations, such as sparse or low-rank subspace clustering, by replacing the norm in Eq. (4) while keeping the same regularizer.
- Because the regularizer is built from raw client data before training, a testable extension is to measure how sensitive FDSC is to the choice of $k$ in k-NN and to noisy or corrupted adjacency graphs.
- If the undefined term in Eq. (8) is resolved by a norm, the method's actual contribution may reduce to a graph-regularized deep subspace clustering objective; comparing against that centralized baseline would isolate the benefit of federation from the benefit of regularization.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes FDSC, a federated deep subspace clustering method in which each client trains a local deep subspace clustering network (shared encoder, private self-expressive layer, private decoder) and the encoder is periodically averaged at a central server. The method adds a regularizer intended to align the self-expressive matrix with a k-NN adjacency matrix, motivated by preserving local neighborhood structure. The authors evaluate FDSC on MNIST, ORL, COIL20, and COIL100, comparing against centralized methods LRSC, DLRSC, and DSCN, and report improvements in ACC, NMI, AMI, and ARI. They also provide t-SNE visualizations and self-expression matrix visualizations to support the claim that federated training improves subspace clustering.
Significance. If the proposed method were correctly specified and the empirical comparisons were rigorous, the paper would address a real gap: extending deep subspace clustering to the federated setting, where clients cannot share raw data. The paper also proposes a potentially useful architectural choice, keeping the self-expressive layer and decoder local while sharing only the encoder. However, the central empirical claim is unsupported because the local training objective in Eq. (8) is undefined as written, so the FDSC2 results that drive the main comparison cannot be interpreted or reproduced. The paper also lacks federated baselines and reports no variance or repeated-run information, further weakening the empirical conclusions. Because the undefined objective is load-bearing for every reported result, the contribution cannot be accepted in its current form.
major comments (3)
- [§4.5, Table 2] The client objective in Eq. (8) is not a scalar function. The first three terms are scalars: (1/2)||X_i - X_hat_i||_F^2, λ1||R_i||_p, and (λ2/2)||Z_i - Z_i R_i||_F^2. The added term λ3(αA_i - βR_i) is an n×n matrix, because A_i is the k-NN adjacency matrix and R_i is the self-expressive matrix. Adding a matrix to scalars is undefined, so no scalar loss exists for local training. This is not a typographical detail: Section 3.4 explicitly states that the algorithm 'compute[s] λ3(αA_i−βR_i)' and stores A_i before training, Algorithm 1 trains using Eq. (8), and Section 4.5 defines FDSC2 by setting λ3=1e6 and attributes FDSC2's improvements over FDSC1 to this term. Consequently, every FDSC2 result in Tables 2–4, the FDSC1-vs-FDSC2 ablation, and the visualization comparisons rest on an undefined objective. Even if one conjectures that a matrix norm is missing, the paper does not specify the corrected objective, so the experiments cannot be reproduced as written.
- [§4.2 and §4.6] The claim that FDSC outperforms state-of-the-art methods is not supported by the experimental design because Table 2 compares FDSC only against centralized subspace clustering methods (LRSC, DLRSC, DSCN), even though the related work in Section 2.3 cites federated clustering methods such as k-FED and FFCM. Without experiments against federated clustering baselines, the paper's statement that FDSC is the first federated clustering model using deep subspace clustering, and the broader claim of superiority, cannot be evaluated. A fair comparison requires at least k-FED, FFCM, or a federated auto-encoder baseline with local k-means/spectral clustering.
- [§4.5, Table 2] The empirical evaluation lacks statistical support. Hyperparameters (λ1, λ2, λ3, m, r, τ, q, batch size) are selected per dataset with no validation procedure, and every reported result is a single number without error bars, standard deviations, or repeated runs. Tables 3 and 4 additionally vary the number of clients and batch size on the same test data, which makes it impossible to distinguish genuine improvements from noise or from overfitting to the test set. Consequently, the observed improvements of roughly 1–2 percentage points in Table 2 are not established as significant.
minor comments (5)
- [§2.1, Eq. (1)] The text defines ||·||_F as the 'Fibonacci norm'; this appears to be a typo for the Frobenius norm, used throughout the paper.
- [§3.2, Eq. (5)] The objective in Eq. (5) is written as min over E* of a sum involving min over E_i, R_i, D_i, but f_i does not depend on E* directly; the relationship between the global encoder E* and the local encoder E_i should be clarified.
- [§4.2] The sentence 'We randomly selected samples with q categories into these subsets' is ambiguous. It should state whether each client receives data from exactly q classes or whether q classes are sampled and then distributed randomly, as this is central to the heterogeneous-data setup.
- [Figures 3 and 4] The figure captions and text use both 'FDSC' and 'FSDC' (e.g., 'using FSDC' in Section 4.3); the consistent acronym should be used.
- [§3.4] In Section 3.4, 'the sever is responsible for communication' contains a typo: 'sever' should be 'server'.
Circularity Check
No circularity: FDSC's effectiveness claim is an empirical comparison against external baselines; the undefined matrix term in Eq. (8) is a correctness flaw, not a derivation loop.
full rationale
The paper makes no first-principles derivation whose conclusion is equivalent to its inputs. The local objective in Eq. (4) is the standard deep subspace clustering loss, and Eq. (8) augments it with an explicit adjacency-alignment regularizer; neither defines the evaluation metric in terms of the optimized variable. The federated encoder update in Eqs. (5)-(6) is FedAvg-style averaging and does not smuggle in the clustering result. The reported ACC/NMI/AMI/ARI values are empirical comparisons against external baselines (LRSC, DLRSC, DSCN), not quantities fitted and then renamed as predictions. Hyperparameters are fixed in Section 4.2 rather than fitted to a subset and predicted on a closely related quantity. The only self-citation, reference [44] by the first author, appears in the introduction as an example of dimensionality reduction and is not load-bearing. The serious flaw in the paper is that Eq. (8) adds the n-by-n matrix term λ3(αA_i−βR_i) to scalar loss terms, making the local objective undefined as written; this undermines the reproducibility and validity of the FDSC2 experimental results, but it is an internal inconsistency rather than a circular reduction of a derived claim to its inputs. Therefore no circularity step can be exhibited under the stated rules.
Assumptions & free parameters
free parameters (6)
- lambda1 =
1 (MNIST, COIL20, COIL100), 2 (ORL)
- lambda2 =
15 (MNIST), 0.2 (ORL), 75 (COIL20), 15 (COIL100)
- lambda3 =
0 (FDSC1) or 1e6 (FDSC2)
- alpha, beta =
alpha=1, beta=1
- k in k-NN adjacency =
not reported
- local epochs tau =
7
assumptions (4)
- domain assumption Self-expressive property holds in the latent representation Z_i, i.e., each point is a linear combination of other points in the same subspace.
- domain assumption Averaging client encoders via Eq. (6) yields a shared encoder useful for all clients' local clustering.
- domain assumption The k-NN adjacency matrix of raw data X_i is a valid target for the self-expressive matrix R_i.
- ad hoc to paper The loss in Eq. (8) is a well-defined scalar objective.
Cite this review
Pith. "Pith review of Federated Deep Subspace Clustering." pith.science (2026). https://pith.science/paper/MFGTDWXF
@misc{pith2026250100230,
author = {Pith},
title = {Pith review of: Federated Deep Subspace Clustering},
year = {2026},
howpublished = {\url{https://pith.science/paper/MFGTDWXF}},
note = {Machine review of arXiv:2501.00230}
}
read the original abstract
This paper introduces FDSC, a private-protected subspace clustering (SC) approach with federated learning (FC) schema. In each client, there is a deep subspace clustering network accounting for grouping the isolated data, composed of a encode network, a self-expressive layer, and a decode network. FDSC is achieved by uploading the encode network to communicate with other clients in the server. Besides, FDSC is also enhanced by preserving the local neighborhood relationship in each client. With the effects of federated learning and locality preservation, the learned data features from the encoder are boosted so as to enhance the self-expressiveness learning and result in better clustering performance. Experiments test FDSC on public datasets and compare with other clustering methods, demonstrating the effectiveness of FDSC.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[1]
Mahdi Abavisani and Vishal M Patel. 2018. Deep multimodal subspace clustering networks. IEEE Journal of Selected Topics in Signal Processing 12, 6 (2018), 1601– 1614
work page 2018
-
[2]
KS Arikumar, Sahaya Beni Prathiba, Mamoun Alazab, Thippa Reddy Gadekallu, Sharnil Pandya, Javed Masood Khan, and Rajalakshmi Shenbaga Moorthy. 2022. FL-PMI: federated learning-based person movement identification through wear- able devices in smart healthcare systems. Sensors 22, 4 (2022), 1377
work page 2022
-
[3]
Sangwon Baek, Gangjoon Yoon, Jinjoo Song, and Sang Min Yoon. 2021. Deep self-representative subspace clustering network. Pattern Recognition 118 (2021), 108041
work page 2021
-
[4]
Syreen Banabilah, Moayad Aloqaily, Eitaa Alsayed, Nida Malik, and Yaser Jarar- weh. 2022. Federated learning review: Fundamentals, enabling technologies, and future applications. Information processing & management 59, 6 (2022), 103061
work page 2022
-
[5]
Kamal Berahmand, Mehrnoush Mohammadi, Azadeh Faroughi, and Rojiar Pir Mohammadiani. 2022. A novel method of spectral clustering in attributed net- works by constructing parameter-free affinity matrix. Cluster Computing (2022), 1–20
work page 2022
-
[6]
Ying Chen, Chun-Guang Li, and Chong You. 2020. Stochastic sparse subspace clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4155–4164
work page 2020
-
[7]
Yongyong Chen, Shuqin Wang, Chong Peng, Zhongyun Hua, and Yicong Zhou
-
[8]
Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. 2021. Exploiting shared representations for personalized federated learning. In Inter- national Conference on Machine Learning . PMLR, 2089–2099
work page 2021
Show all 48 references
-
[9]
Don Kurian Dennis, Tian Li, and Virginia Smith. 2021. Heterogeneity for the win: One-shot federated clustering. In International Conference on Machine Learning . PMLR, 2611–2620
2021
-
[10]
Ehsan Elhamifar and René Vidal. 2013. Sparse subspace clustering: Algorithm, theory, and applications. IEEE transactions on pattern analysis and machine intelligence 35, 11 (2013), 2765–2781
2013
-
[11]
Xiuwen Fang and Mang Ye. 2022. Robust federated learning with noisy and heterogeneous clients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 10072–10081
2022
-
[12]
Dashan Gao, Ben Tan, Ce Ju, Vincent W Zheng, and Qiang Yang. 2020. Privacy threats against federated matrix factorization. arXiv preprint arXiv:2007.01587 (2020)
2020 arXiv
-
[13]
Avishek Ghosh, Jichan Chung, Dong Yin, and Kannan Ramchandran. 2020. An efficient framework for clustered federated learning. Advances in Neural Infor- mation Processing Systems 33 (2020), 19586–19597
2020
-
[14]
Benjamin D Haeffele, Chong You, and René Vidal. 2020. A critique of self- expressive deep subspace clustering. arXiv preprint arXiv:2010.03697 (2020)
2020 arXiv
-
[15]
Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, Chuhan Wu, Xing Xie, and Meeyoung Cha. 2022. FedX: Unsupervised Federated Learning with Cross Knowledge Distillation. In Computer Vision–ECCV 2022: 17th European Conference, Tel A viv, Israel, October 23–27, 2022, Proceeding...
2022
-
[16]
Pan Ji, Mathieu Salzmann, and Hongdong Li. 2014. Efficient dense subspace clustering. In IEEE Winter conference on applications of computer vision . IEEE, 461–468
2014
-
[17]
Pan Ji, Tong Zhang, Hongdong Li, Mathieu Salzmann, and Ian Reid. 2017. Deep subspace clustering networks. Advances in neural information processing systems 30 (2017)
2017
-
[18]
Mohsen Kheirandishfard, Fariba Zohrizadeh, and Farhad Kamangar. 2020. Deep low-rank subspace clustering. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition workshops . 864–865
2020
-
[19]
Jianjun Lei, Xinyu Li, Bo Peng, Leyuan Fang, Nam Ling, and Qingming Huang
-
[20]
Kun Li, Yao Qin, Qiang Ling, Yingqian Wang, Zaiping Lin, and Wei An. 2021. Self-supervised deep subspace clustering for hyperspectral images with adaptive self-expressive coefficient matrix initialization. IEEE Journal of Selected Topics in Applied Earth Observations and Remot...
2021
-
[21]
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 systems 2 (2020), 429–450
2020
-
[22]
Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. 2019. On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189 (2019)
2019 arXiv
-
[23]
Zengyi Li, Yubei Chen, Yann LeCun, and Friedrich T Sommer. 2022. Neural manifold clustering and embedding. arXiv preprint arXiv:2201.10000 (2022)
2022 arXiv
-
[24]
Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer us- ing shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision. 10012–10022
2021
-
[25]
Guodong Long, Ming Xie, Tao Shen, Tianyi Zhou, Xianzhi Wang, and Jing Jiang
-
[26]
Ekdeep Singh Lubana, Chi Ian Tang, Fahim Kawsar, Robert P Dick, and Akhil Mathur. 2022. Orchestra: Unsupervised federated learning via globally consistent clustering. arXiv preprint arXiv:2205.11506 (2022)
2022 arXiv
-
[27]
Juncheng Lv, Zhao Kang, Xiao Lu, and Zenglin Xu. 2021. Pseudo-supervised deep subspace clustering. IEEE Transactions on Image Processing 30 (2021), 5252–5263
2021
-
[28]
Yishay Mansour, Mehryar Mohri, Jae Ro, and Ananda Theertha Suresh. 2020. Three approaches for personalization with applications to federated learning. arXiv preprint arXiv:2002.10619 (2020)
2020 arXiv
-
[29]
Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-efficient learning of deep net- works from decentralized data. In The Proceedings of Artificial intelligence and statistics. PMLR, 1273–1282
2017
-
[30]
Andrew Ng, Michael Jordan, and Yair Weiss. 2001. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems 14 (2001)
2001
-
[31]
Dinh C Nguyen, Ming Ding, Pubudu N Pathirana, Aruna Seneviratne, Jun Li, and H Vincent Poor. 2021. Federated learning for internet of things: A comprehensive survey. IEEE Communications Surveys & Tutorials 23, 3 (2021), 1622–1658
2021
-
[32]
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
-
[33]
Xi Peng, Jiashi Feng, Joey Tianyi Zhou, Yingjie Lei, and Shuicheng Yan. 2020. Deep subspace clustering. IEEE transactions on neural networks and learning systems 31, 12 (2020), 5509–5521
2020
-
[34]
Leif E Peterson. 2009. K-nearest neighbor. Scholarpedia 4, 2 (2009), 1883
2009
-
[35]
Felix Sattler, Klaus-Robert Müller, and Wojciech Samek. 2020. Clustered feder- ated learning: Model-agnostic distributed multitask optimization under privacy constraints. IEEE transactions on neural networks and learning systems 32, 8 (2020), 3710–3722
2020
-
[36]
Morris Stallmann and Anna Wilbik. 2022. Towards Federated Clustering: A Federated Fuzzy𝑐-Means Algorithm (FFCM). arXiv preprint arXiv:2201.07316 (2022)
2022 arXiv
-
[37]
Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. 2022. Fedproto: Federated prototype learning across hetero- geneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 36. 8432–8440
2022
-
[38]
Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)
2008
-
[39]
René Vidal and Paolo Favaro. 2014. Low rank subspace clustering (LRSC).Pattern Recognition Letters 43 (2014), 47–61
2014
-
[40]
Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. 2019. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST) 10, 2 (2019), 1–19
2019
-
[41]
Shuai Yang, Wenqi Zhu, and Yuesheng Zhu. 2020. Residual encoder-decoder network for deep subspace clustering. In 2020 IEEE International Conference on Image Processing (ICIP). IEEE, 2895–2899
2020
-
[42]
Hongyan Zhang, Han Zhai, Liangpei Zhang, and Pingxiang Li. 2016. Spectral– spatial sparse subspace clustering for hyperspectral remote sensing images. IEEE Transactions on Geoscience and Remote Sensing 54, 6 (2016), 3672–3684
2016
-
[43]
Qingchen Zhang, Laurence T Yang, Zhikui Chen, and Peng Li. 2018. A survey on deep learning for big data. Information Fusion 42 (2018), 146–157
2018
-
[44]
Yupei Zhang, Ming Xiang, and Bo Yang. 2017. Low-rank preserving embedding. Pattern Recognition 70 (2017), 112–125
2017
-
[45]
Pengfei Zhu, Binyuan Hui, Changqing Zhang, Dawei Du, Longyin Wen, and Qinghua Hu. 2019. Multi-view deep subspace clustering networks. arXiv preprint arXiv:1908.01978 (2019)
2019 arXiv
-
[2020]
IEEE Transactions on Circuits and Systems for Video Technology 31, 7 (2020), 2686–2697
Deep spatial-spectral subspace clustering for hyperspectral image. IEEE Transactions on Circuits and Systems for Video Technology 31, 7 (2020), 2686–2697
2020
-
[2021]
IEEE Transactions on Image Processing 30 (2021), 4022–4035
Generalized nonconvex low-rank tensor approximation for multi-view subspace clustering. IEEE Transactions on Image Processing 30 (2021), 4022–4035
2021
-
[2023]
World Wide Web 26, 1 (2023), 481–500
Multi-center federated learning: clients clustering for better personalization. World Wide Web 26, 1 (2023), 481–500
2023
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.