GraMFedDHAR: Graph Based Multimodal Differentially Private Federated HAR

Labani Halder; Sarbani Palit; Tanmay Sen

REVIEW 5 major objections 5 minor 22 references

GraMFedDHAR: Graph Based Multimodal Differentially Private Federated HAR

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Attention-fused graph networks lose 7–13 percentage points less accuracy than feedforward networks under differential-privacy noise in multimodal activity recognition.

desk verdict GCNs under DP noise is a real empirical signal, but the missing privacy accounting makes the headline numbers unverifiable until the authors clarify how epsilon was computed. read the letter →

arxiv 2509.05671 v1 pith:FS2HHVTQ submitted 2025-09-06 cs.LG cs.AIcs.CRstat.ML

Labani Halder , Tanmay Sen , Sarbani Palit This is my paper · ORCID

classification cs.LGcs.AIcs.CRstat.ML

keywords humanactivityrecognitionmultimodalfederatedlearninggraphconvolutionalnetworksdifferentialprivacyDP-SGDattention-basedfusionsensortimeseriesprivacy-utilitytradeoff

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 GraMFedDHAR, a federated framework for multimodal human activity recognition in which each sensor stream is turned into a graph, processed by residual graph convolutional layers, and fused by attention. Its central empirical claim is that this graph architecture loses far less accuracy than a comparable feedforward network when differential-privacy noise is added: up to 2 percentage points better without privacy, and 7–13 percentage points better under privacy constraints, in both centralized and federated training. If true, the architectural choice matters for privacy: graph-based message passing absorbs DP noise better than concatenation-style feedforward fusion. The paper supports the claim with accuracy and F1 tables, confusion matrices, t-SNE embeddings, convergence curves, and ablations over modality combinations and client participation.

What carries the argument

Modality-specific graphs are the central object: each time window becomes a node, edges link consecutive windows to capture temporal dependence, and a distance-based adjacency captures spatial similarity across sensor readings. Residual GCN layers propagate features within each modality, and an attention mechanism computes per-node modality weights before fusing embeddings into one classification vector. Privacy is enforced by DP-SGD: client gradients are clipped and augmented with Gaussian noise before FedAvg aggregation.

What would settle it

Recompute the actual privacy budget for Table 2 using a documented standard accountant with the stated σ, δ, client sampling fraction, batch size, 50 federated rounds, and 20 local epochs; then replot GCN-versus-FFN accuracy at matched ε values. If the gap is not 7–13 percentage points at equal privacy, the central claim fails. A second check: replace the GCN's message passing with mean-pooling inside the feedforward model; if the gap disappears, the graph structure itself is the source of robustness.

Watch

Extended reading notes

Core claim

On the MEx physiotherapy dataset, using two accelerometers, a depth camera, and a pressure mat, the paper compares a multimodal attention-fused GCN (MultiModalGCN) with a same-sized feedforward network (MultiModalFFN). Without DP, the GCN is slightly ahead, reaching 0.9878 vs 0.9650 centralized and 0.9941 vs 0.9850 federated. Under strong DP (ε = 0.5), the GCN retains 0.8075 vs 0.6708 centralized and 0.8917 vs 0.8126 federated; under weaker DP (ε = 1.5), it reaches 0.9011 vs 0.7308 centralized and 0.9562 vs 0.8863 federated. The paper interprets these gaps as evidence that graph-based modeling is more resilient to the utility loss introduced by differential privacy, and that federated traini

Load-bearing premise

The reported ε values are correctly computed from the added noise and the number of training steps; if the privacy accounting is wrong, the graph model is being compared at a different privacy strength than the feedforward model, and the claimed 7–13 percentage-point advantage may not hold at equal privacy.

Editorial extensions

If this is right

Deploying graph-based fusion instead of plain concatenation gives a privacy-preserving HAR system the same formal privacy guarantee with measurably less accuracy loss.
Under strict privacy budgets, federated training is the safer paradigm: the paper finds it degrades less than centralized training at ε = 0.5.
When privacy is tight, increasing the fraction of clients sampled per round is a concrete lever for recovering accuracy, because averaging cancels DP noise.
Multimodality becomes more valuable under DP: ablations show multimodal combinations hold up better than single modalities, especially in federated settings.

Reading between the lines

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

Editorial inference: If the reported privacy budgets are verified, the same graph-versus-feedforward advantage may generalize to other DP multimodal tasks, because graph message passing acts as a structured averaging that partially denoises the added Gaussian noise.
Editorial inference: The paper never states its privacy accountant, so the 7–13 percentage-point gap is currently a comparison at stated noise scales; an independent recomputation of ε is needed before treating the comparison as equal-privacy evidence.
Editorial inference: A testable extension is to add graph-style smoothing or message passing to the feedforward baseline; if the gap closes, the source of robustness is the graph's averaging structure rather than attention fusion or extra parameters.

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.

Desk Editor's Note

read the letter

The one genuinely new thing here is the empirical comparison: on the MEx dataset, a multimodal graph model (MultiModalGCN) degrades less than a feedforward baseline under DP noise, with gaps of 7–13% in accuracy across centralized and federated settings. That finding is not in the cited literature, and if it holds, it is a useful data point for the HAR community. The paper also deserves credit for running a fairly thorough ablation suite—modality combinations, client participation, utility loss, t-SNE, and confusion matrices. The consistency of the trend across those tables is the strongest evidence in favor of the claim.

The soft spots are the usual suspects, but they bite harder here because the central claim is about behavior under differential privacy. The DP accounting is simply not verifiable: no accountant, no total number of composition steps, no clear statement of what the “sampling rate for DP” (0.01) actually is. The internal inconsistency between ε=1.5 and ε=2.0 across tables and figures is not cosmetic—it suggests the epsilon values were assigned, not computed. If the two architectures were actually compared at different true privacy levels, the 7–13% gap could just be a noise-level artifact. That is exactly the kind of load-bearing flaw that a referee needs to flag. The convergence guarantee in Section 5.4 is also borrowed from a bound that assumes strong convexity; GCNs with cross-entropy loss do not satisfy that, so the theory is decorative rather than supportive. Minor issues: no error bars or repeated runs, and the graph construction/fusion details are underspecified enough that reproduction would be hard. I did not find the “first to apply” framing or the self-citation to be a real problem; the novelty is modest but legitimate.

Who is this for? Researchers building privacy-preserving HAR systems, especially anyone deciding between GCNs and feedforward backbones. The paper deserves a serious referee because the empirical direction is plausible and potentially influential, but it needs major revision before publication.

My recommendation: send it to review, but ask the authors to provide a concrete privacy accounting (RDP or moments accountant, with T and q specified), align all reported epsilons, add error bars, and soften the convergence claims. If they cannot produce the accounting, the paper should not be accepted.

Referee Report

5 major / 5 minor

Summary. The paper proposes GraMFedDHAR, a federated learning framework for multimodal human activity recognition (HAR). Each sensor stream is converted into a graph via temporal windows and distance-based edges, encoded by modality-specific GCNs, fused by attention, and classified. Differential privacy is applied by clipping client gradients and adding Gaussian noise before FedAvg. Experiments on the MEx dataset compare MultiModalGCN with a feedforward baseline in centralized and federated settings, with and without DP. The headline claim is that GCNs are more resilient to DP-induced utility loss, with 7–13% accuracy gaps under DP. The paper also provides a convergence bound and ablations over modality subsets and client participation.

Significance. If the empirical claim is correct, the paper offers a useful architectural insight: graph-structured multimodal representations may degrade more gracefully under DP noise than concatenative feedforward baselines, and federated training may mitigate DP damage compared with centralized training. The paper uses a public dataset, compares with a matched feedforward baseline, and provides extensive ablations (modality combinations, client participation, privacy levels). These are strengths. However, the central claim is not currently verifiable: the DP budgets in Tables 2–6 are not backed by a privacy accountant or composition count, the weak-DP label is internally inconsistent (ε=1.5 vs ε=2.0), and the reported client sampling rate would select zero clients. The convergence guarantee in §5.4 rests on strong convexity, which GCNs do not satisfy. Without fixing these load-bearing points, the paper's main contribution—equal-privacy superiority of GCNs—is unsubstantiated.

major comments (5)

[§6.2.3, Table 2, Table 1] The DP accounting is not verifiable. Table 2 reports (ε, σ, δ) but no privacy accountant (RDP? moments accountant?), no total number of DP gradient steps or composition count, and no explanation of how 'Sampling rate for DP' in Table 1 enters the accounting. If this is the client sampling fraction q of Algorithm 1, then with 30 clients q=0.01 gives floor(0.3)=0 clients per round, making the protocol impossible as written. If it is a per-example Poisson sampling rate, the mini-batching, rounds, and composition are not specified. Moreover, the weak-DP setting is internally inconsistent: Table 2 and §6.3.1 use ε=1.5, σ=0.8629, while Figures 1, 5, 6 and Tables 4–6 use ε=2.0, σ=0.7852. Without a consistent, valid accountant, the headline 7–13% gaps may reflect different noise levels rather than GCN robustness under equal privacy.
[Algorithm 1 / Table 1] Algorithm 1 cannot execute as written: line 3 selects a subset of size floor(q|C|), with q=0.01 and |C|=30 this is 0 clients. The paper must clarify the client sampling mechanism, the number of clients actually used, and how the 'sampling rate for DP' relates to both the FedAvg protocol and the DP composition. As written, the central algorithm is not reproducible.
[§5.4, Eq. (1)] The convergence bound in §5.4 assumes F is L-smooth and μ-strongly convex (Assumptions 1–2 in §5.3). A GCN with ReLU activations and cross-entropy loss is nonconvex, so the linear convergence bound, borrowed from [22], does not apply. No derivation is given for the DP noise term dσ²C²/m either. The theoretical claim should either be removed or replaced by an analysis appropriate for nonconvex objectives; as it stands, it does not support Algorithm 1.
[Tables 2, 4–6] All numerical results are single-run point estimates with no error bars, standard deviations, or number of seeds. DP-SGD noise and client sampling make the results stochastic; without variance information, the 7–13% DP gaps and even the 1–2% non-DP gaps cannot be assessed for statistical significance. The authors should report mean ± std over at least 5 independent runs and specify the random seeds.
[§6.2.3, Table 1] The centralized and federated settings are not matched: centralized uses 500 epochs and learning rate 0.001, while federated uses 50 rounds × 20 local epochs = 1000 local client epochs and learning rate 0.01. The comparison 'centralized vs federated' is therefore confounded by computation budget and hyperparameters. Additionally, Algorithm 1 performs a single gradient step per client per round, but §6.2.3 states Adam is used; the algorithm and implementation must be reconciled.

minor comments (5)

[§6.3.1, §6.2.3, §6.2.4] Typographical and grammatical errors: 'subtantial' (§6.3.1), 'tranining' (§6.2.3), 'modals' (§6.2.4), 'central sever' (§3.2). A careful proofread is needed.
[§6.3.5, Table 3] The text says 'Table III presents boxplots', but Table 3 appears to be a numeric table with a caption about utility loss distributions. Either reference a figure or update the caption to match the content.
[Abstract, §6.2.1] The model is called a 'residual GCN' in the abstract and conclusion, but the architecture description in §6.2.1 lists GCN layers with normalization, ReLU, and dropout, without explicit residual connections. Please either specify the residual connections or remove the term.
[Algorithm 1, §5.2] Line 7 sends Δw_c = w_c − w_global, while line 9 averages these updates. It should be clarified that the update is the clipped noisy gradient, not a raw weight difference, and whether the local step is full-batch or mini-batch.
[Table 2, Figures 3–6] The notation 'ε=∞' for no-DP is unconventional; consider labeling as 'No DP'. Also verify the axis labels of Figures 5 and 6: the text describes accuracy but the captions suggest 'F1 Score Loss'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: reported accuracies are direct experimental measurements; the convergence bound is an external citation not used to predict results, and self-citations are background only.

full rationale

The paper's central claims are empirical: MultiModalGCN is compared against MultiModalFFN on the MEx dataset under centralized/federated settings and with/without DP (Abstract, Table 2, Figures 3-8). These accuracy/F1 numbers are direct measurements from runs, not values derived from a fitted equation or from the convergence guarantee. The convergence bound in Section 5.4 is quoted from Rizk and Sayed [22], an external source, and it is not used to generate the table entries; it is a standalone theoretical statement. No parameter is fitted to a subset and then presented as a prediction. The graph construction, attention fusion, and DP-SGD procedure are defined by the paper and evaluated empirically, so there is no self-definitional reduction. The two self-citations are to prior related work by one co-author: [10] MHARFedLLM and [13] GraFeHTy. Both are used only as related-work context and to motivate novelty; neither is invoked as a load-bearing uniqueness theorem, nor as justification for the main performance comparison. The skeptical concern about DP accounting (no privacy accountant, no explicit composition count, inconsistent ε values of 1.5 vs 2.0 across Table 2 and Figures 1/5/6) is a correctness/verifiability issue, not a circularity issue: the ε labels are experimental hyperparameters reported as conditions, and an incorrect or unverifiable ε does not make the derivation circular. The utility-loss metric is defined as a post-hoc ratio of measured accuracies, not fitted to make a point. Overall, the derivation chain is self-contained in the sense that the findings rest on direct experiment rather than on assumptions that already contain the conclusions.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new physical entities. It relies on several tuned hyperparameters and on the assumption that the reported DP parameters correspond to a valid privacy accounting, which the paper does not demonstrate.

free parameters (6)

DP noise scale sigma = 1.4216 (epsilon=0.5), 0.8629 (epsilon=1.5), 0.7852 (epsilon=2.0)
Stated in Section 6.3.1 and figure captions; chosen to hit target privacy budgets, but the mapping from (sigma, sampling, steps) to epsilon is not shown.
Data sampling rate for DP = 0.01
Listed in Table 1; ambiguous whether data-level sampling or client-level sampling; directly affects privacy composition and is not explained.
Client sampling fraction q = Not reported in main table
Used in ablation (Figure 7) and in Algorithm 1 (client sampling), but the value for the main results is missing; affects both convergence and DP amplification.
Graph edge threshold = Not reported
Section 6.2.1 says edges are created 'based on the similarity distance function using threshold based tuning'; no threshold value or tuning procedure is given.
DC/PM feature autoencoder latent dimension = 64
Fully connected autoencoder with layers 512-256-64 is trained on depth camera and pressure mat data; the latent size is chosen without ablation.
DCT retained coefficients = 60 per axis
Discrete Cosine Transformation applied per axis for accelerometer data, retaining 60 coefficients per axis, yielding 180-dimensional vectors; chosen without stated rationale.

assumptions (4)

ad hoc to paper The global loss F(w) is L-smooth and mu-strongly convex
Section 5.4 states these as assumptions to invoke the convergence bound. GCNs with cross-entropy loss are not strongly convex, so the bound does not apply to the actual model.
domain assumption Gradients are bounded by G
Assumption 3 in Section 5.4; standard in DP-FL analysis, but not verified for the GCN.
domain assumption The graph topology constructed from distance-based similarity captures temporal and cross-modal dependencies relevant to HAR
Section 5.1 and 6.2.1 describe constructing time-window nodes and threshold-based edges; whether this graph structure is informative is assumed, not justified.
domain assumption The MEx dataset labels and sensor alignments are correct
Standard experimental assumption; the paper relies on the dataset's ground truth.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GraMFedDHAR: Graph Based Multimodal Differentially Private Federated HAR." pith.science (2026). https://pith.science/paper/FS2HHVTQ

@misc{pith2026250905671,
  author       = {Pith},
  title        = {Pith review of: GraMFedDHAR: Graph Based Multimodal Differentially Private Federated HAR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FS2HHVTQ}},
  note         = {Machine review of arXiv:2509.05671}
}

read the original abstract

Human Activity Recognition (HAR) using multimodal sensor data remains challenging due to noisy or incomplete measurements, scarcity of labeled examples, and privacy concerns. Traditional centralized deep learning approaches are often constrained by infrastructure availability, network latency, and data sharing restrictions. While federated learning (FL) addresses privacy by training models locally and sharing only model parameters, it still has to tackle issues arising from the use of heterogeneous multimodal data and differential privacy requirements. In this article, a Graph-based Multimodal Federated Learning framework, GraMFedDHAR, is proposed for HAR tasks. Diverse sensor streams such as a pressure mat, depth camera, and multiple accelerometers are modeled as modality-specific graphs, processed through residual Graph Convolutional Neural Networks (GCNs), and fused via attention-based weighting rather than simple concatenation. The fused embeddings enable robust activity classification, while differential privacy safeguards data during federated aggregation. Experimental results show that the proposed MultiModalGCN model outperforms the baseline MultiModalFFN, with up to 2 percent higher accuracy in non-DP settings in both centralized and federated paradigms. More importantly, significant improvements are observed under differential privacy constraints: MultiModalGCN consistently surpasses MultiModalFFN, with performance gaps ranging from 7 to 13 percent depending on the privacy budget and setting. These results highlight the robustness of graph-based modeling in multimodal learning, where GNNs prove more resilient to the performance degradation introduced by DP noise.

Figures

Figures reproduced from arXiv: 2509.05671 by the authors.

**Figure 1.** t-SNE visualization under Weak DP (ε = 2.0, σ = 0.7852) for MultiModalFFN and MultiModalGCN in centralized and federated paradigms. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

**Figure 2.** t-SNE visualization under Strong DP (ε = 0.5, σ = 1.4216) for MultiModalFFN and MultiModalGCN in centralized and federated paradigms. Figure (1) and (2) present the t-SNE projections of fused multi modal embeddings (act, acw, dc, pm) learned under both weak and strong DP settings. In the centralized setting, both model suffer from noise-induced degradation. The centralized FFN shows significant class overlap and poo… view at source ↗

**Figure 3.** Impact of privacy level (ε) on model convergence in Centralized Learning. 6.3.4 Comparative Evaluation: Graph-based vs FeedForward Architectures with DP Figures (5) and (6) provide a comparative analysis of graph-based and feedforward models under strong and weak DP cases. In centralized learning, GCN consistently outperforms than FFN. Under federated paradigm, FFN shows a slight early advantage, surpassing GCN only… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Impact of privacy level (ε) on model convergence in Federated Learning. 6.3.7 Ablation study of Modality Combinations Across Privacy Levels and Learning setting From Table (4), (5), and (6) multi-modal combinations significantly boost performance, especially in federat…

**Figure 5.** Figure 5: Performance metrics under Weak DP (ε = 2.0, σ = 0.7852) for centralized and federated models [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: Performance metrics under Strong DP (ε = 0.5, σ = 1.4216) for centralized and federated models [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Effect of client participation on accuracy under DP. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

**Figure 8.** Figure 8: Confusion matrices under varying privacy budgets ( [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages

[22]

Elsa Rizk and Ali H. Sayed. A graph federated architecture with privacy preserving learning.CoRR, abs/2104.13215, 2021. 15

work page Pith review arXiv 2021
[1]

Mmhar-ensemnet: a multi-modal human activity recognition model.IEEE Sensors Journal, 21(10):11569–11576, 2020

Avigyan Das, Pritam Sil, Pawan Kumar Singh, Vikrant Bhateja, and Ram Sarkar. Mmhar-ensemnet: a multi-modal human activity recognition model.IEEE Sensors Journal, 21(10):11569–11576, 2020

work page 2020
[2]

A unified framework for multi-modal federated learning.Neurocomputing, 480:110–118, 2022

Baochen Xiong, Xiaoshan Yang, Fan Qi, and Changsheng Xu. A unified framework for multi-modal federated learning.Neurocomputing, 480:110–118, 2022

work page 2022
[3]

Cross-modal federated human activity recogni- tion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5345–5361, 2024

Xiaoshan Yang, Baochen Xiong, Yi Huang, and Changsheng Xu. Cross-modal federated human activity recogni- tion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5345–5361, 2024. 14 APREPRINT- SEPTEMBER9, 2025

work page 2024
[4]

Differentially private human activity recognition for smartphone users.Multimedia Tools and Applications, 81(28):40827–40848, 2022

Avishek Garain, Rudrajit Dawn, Saswat Singh, and Chandreyee Chowdhury. Differentially private human activity recognition for smartphone users.Multimedia Tools and Applications, 81(28):40827–40848, 2022

work page 2022
[5]

Deep learning with differential privacy

Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. InProceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308–318, 2016

2016
[6]

Differential privacy for deep and federated learning: A survey.IEEE access, 10:22359–22380, 2022

Ahmed El Ouadrhiri and Ahmed Abdelhadi. Differential privacy for deep and federated learning: A survey.IEEE access, 10:22359–22380, 2022

work page 2022
[7]

Federated learning with differential privacy: Algorithms and performance analysis.IEEE transactions on information forensics and security, 15:3454–3469, 2020

Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H Yang, Farhad Farokhi, Shi Jin, Tony QS Quek, and H Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis.IEEE transactions on information forensics and security, 15:3454–3469, 2020

work page 2020

Show all 22 references

[8]

Temporal differential privacy for human activity recognition

Debaditya Roy and Šar¯unas Girdzijauskas. Temporal differential privacy for human activity recognition. In2023 IEEE 10th International Conference on Data Science and Advanced Analytics (DSAA), pages 1–10. IEEE, 2023

2023
[9]

A new framework for smartphone sensor based human activity recognition using graph neural network.IEEE Sensors Journal, 2020

Riktim Mondal, Debadyuti Mukherjee, Pawan Kumar Singh, Vikrant Bhateja, and Ram Sarkar. A new framework for smartphone sensor based human activity recognition using graph neural network.IEEE Sensors Journal, 2020

2020
[10]

Mharfedllm: Multimodal human activity recognition using federated large language model.arXiv preprint arXiv:2508.01701, 2025

Asmit Bandyopadhyay, Rohit Basu, Tanmay Sen, and Swagatam Das. Mharfedllm: Multimodal human activity recognition using federated large language model.arXiv preprint arXiv:2508.01701, 2025

2025 arXiv
[11]

Body rfid skeleton-based human activity recognition using graph convolution neural network.IEEE Transactions on Mobile Computing, 23(6):7301–7317, 2023

Ziyi Wang, Yihong Chen, Hao Zheng, Meng Liu, and Ping Huang. Body rfid skeleton-based human activity recognition using graph convolution neural network.IEEE Transactions on Mobile Computing, 23(6):7301–7317, 2023

2023
[12]

Po Yang, Congmin Yang, Vitaveska Lanfranchi, and Fabio Ciravegna. Activity graph based convolutional neural network for human activity recognition using acceleration and gyroscope data.IEEE Transactions on Industrial Informatics, 18(10):6619–6630, 2022

2022
[13]

Grafehty: Graph neural network using federated learning for human activity recognition

Abhishek Sarkar, Tanmay Sen, and Ashis Kumar Roy. Grafehty: Graph neural network using federated learning for human activity recognition. In2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA), pages 1124–1129. IEEE, 2021

2021
[14]

Graph convolutional neu- ral network for human action recognition: A comprehensive survey.IEEE Transactions on Artificial Intelligence, 2(2):128–145, 2021

Tasweer Ahmad, Lianwen Jin, Xin Zhang, Songxuan Lai, Guozhi Tang, and Luojun Lin. Graph convolutional neu- ral network for human action recognition: A comprehensive survey.IEEE Transactions on Artificial Intelligence, 2(2):128–145, 2021

2021
[15]

A review of machine learning-based human activity recognition for diverse applications.Neural Computing and Applications, 34(21):18289–18324, 2022

Farzana Kulsoom, Sanam Narejo, Zahid Mehmood, Hassan Nazeer Chaudhry, Ayesha Butt, and Ali Kashif Bashir. A review of machine learning-based human activity recognition for diverse applications.Neural Computing and Applications, 34(21):18289–18324, 2022

2022
[16]

Lightweight transformers for human activity recognition on mobile devices.arXiv preprint arXiv:2209.11750, 2022

Sannara Ek, François Portet, and Philippe Lalanda. Lightweight transformers for human activity recognition on mobile devices.arXiv preprint arXiv:2209.11750, 2022

2022 arXiv
[17]

Differentially private integrated decision gradients (idg-dp) for radar-based human activity recognition

Idris Zakariyya, Linda Tran, Kaushik Bhargav Sivangi, Paul Henderson, and Fani Deligianni. Differentially private integrated decision gradients (idg-dp) for radar-based human activity recognition. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pag...

2025
[18]

Human activity recognition using machine learning methods in a smart healthcare environment

Abdulhamit Subasi, Kholoud Khateeb, Tayeb Brahimi, and Akila Sarirete. Human activity recognition using machine learning methods in a smart healthcare environment. InInnovation in health informatics, pages 123–144. Elsevier, 2020

2020
[19]

Deep convlstm with self-attention for human activity decoding using wearable sensors.IEEE Sensors Journal, 21(6):8575–8582, 2020

Satya P Singh, Madan Kumar Sharma, Aimé Lay-Ekuakille, Deepak Gangwar, and Sukrit Gupta. Deep convlstm with self-attention for human activity decoding using wearable sensors.IEEE Sensors Journal, 21(6):8575–8582, 2020

2020
[20]

Attention is all you need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in neural information processing systems, pages 5998–6008, 2017

2017
[21]

Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

2016 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [22]

Elsa Rizk and Ali H. Sayed. A graph federated architecture with privacy preserving learning.CoRR, abs/2104.13215, 2021. 15

work page Pith review arXiv 2021

[2] [1]

Mmhar-ensemnet: a multi-modal human activity recognition model.IEEE Sensors Journal, 21(10):11569–11576, 2020

Avigyan Das, Pritam Sil, Pawan Kumar Singh, Vikrant Bhateja, and Ram Sarkar. Mmhar-ensemnet: a multi-modal human activity recognition model.IEEE Sensors Journal, 21(10):11569–11576, 2020

work page 2020

[3] [2]

A unified framework for multi-modal federated learning.Neurocomputing, 480:110–118, 2022

Baochen Xiong, Xiaoshan Yang, Fan Qi, and Changsheng Xu. A unified framework for multi-modal federated learning.Neurocomputing, 480:110–118, 2022

work page 2022

[4] [3]

Cross-modal federated human activity recogni- tion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5345–5361, 2024

Xiaoshan Yang, Baochen Xiong, Yi Huang, and Changsheng Xu. Cross-modal federated human activity recogni- tion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5345–5361, 2024. 14 APREPRINT- SEPTEMBER9, 2025

work page 2024

[5] [4]

Differentially private human activity recognition for smartphone users.Multimedia Tools and Applications, 81(28):40827–40848, 2022

Avishek Garain, Rudrajit Dawn, Saswat Singh, and Chandreyee Chowdhury. Differentially private human activity recognition for smartphone users.Multimedia Tools and Applications, 81(28):40827–40848, 2022

work page 2022

[6] [5]

Deep learning with differential privacy

Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. InProceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308–318, 2016

2016

[7] [6]

Differential privacy for deep and federated learning: A survey.IEEE access, 10:22359–22380, 2022

Ahmed El Ouadrhiri and Ahmed Abdelhadi. Differential privacy for deep and federated learning: A survey.IEEE access, 10:22359–22380, 2022

work page 2022

[8] [7]

Federated learning with differential privacy: Algorithms and performance analysis.IEEE transactions on information forensics and security, 15:3454–3469, 2020

Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H Yang, Farhad Farokhi, Shi Jin, Tony QS Quek, and H Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis.IEEE transactions on information forensics and security, 15:3454–3469, 2020

work page 2020