Pith. sign in

REVIEW 3 major objections 6 minor 66 references

RoadFed: A Multimodal Federated Learning System for Improving Road Safety

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read RoadFed claims a device-edge-cloud federated system that fuses image and text data can detect road hazards at 96.42% accuracy with 0.0351-second latency and up to 1,000 times lower communication cost, while preserving privacy via local…

desk verdict Solid engineering integration whose privacy proof doesn't cover its own algorithm and whose non-i.i.d. robustness claim is contradicted by its own Figure 8; worth a serious referee for the system, not for the theory. read the letter →

arxiv 2502.09978 v4 pith:IGIFXOLD submitted 2025-02-14 cs.CE

Yachao Yuan , Zhen Yu , Yali Yuan , Xingyu Chen , Yingwen Wu , Thar Baker This is my paper · ORCID
classification cs.CE
keywords roadhazarddetectionfederatedlearningmultimodallocaldifferentialprivacyedge-cloudcomputingintelligenttransportationsystemscommunicationefficiencytripletloss
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

This paper proposes RoadFed, a federated learning system for detecting and alerting road hazards using both images and text from IoT devices. The authors claim that combining a triplet-loss multimodal detector (MRHD), a communication-efficient federated aggregation scheme (MFed), and a dimension-reduction local differential privacy algorithm (MLDP) solves three problems that earlier systems handled separately: single-modality detectors miss textual cues, federated road-damage models are communication-heavy, and high-dimensional data makes local differential privacy too noisy. On a self-collected five-class dataset and the public CrisisMMD dataset, RoadFed is reported to reach 96.42% accuracy and 92.00% accuracy respectively, detect hazards in 0.0351 seconds, and transmit about 0.004 GB, orders of magnitude less than existing edge-cloud baselines. If these results hold, the framework is a practical template for crowdsourced road-safety services that are fast, scalable, and private.

What carries the argument

Three mechanisms carry the argument. First, MRHD pre-trains MobileNetV2-based image and BERT-based text encoders using a triplet loss over image and text anchors, so that inter-class distances grow and intra-class distances shrink across modalities, then fine-tunes a merging block with cross-entropy loss. Second, MFed reduces communication by transmitting only dynamically quantized int8 weight differences, computed as the difference between the updated local model and the previous global model, with the cloud aggregating the received quantized differences and a decaying learning rate to speed convergence on non-i.i.d. data. Third, MLDP applies a random projection using matrices with entries of plus or minus 1/e, followed by Tanh, to shrink the dimension of high-dimensional text and image data before adding Laplace noise at a scale derived from the L1 sensitivity divided by the privacy budget, with the goal of achieving local differential privacy at a lower error rate than perturbing the original high-dimensional input.

What would settle it

Compute the L1 sensitivity of the actual perturbed pipeline, namely the composed function $f(x)=\mathrm{Tanh}(Qx)$ for text and $f(X)=\mathrm{Tanh}(QXR)$ for images, over all pairs of neighboring inputs using the matrices from Eq. (9). If the maximum of $\|f(x)-f(y)\|_1$ exceeds the $s_1(f)$ value used to set the Laplace scale in Algorithm 2, then the epsilon-LDP proof in Eqs. (8)-(13) does not apply to what RoadFed actually deploys, and the privacy pillar fails for that input dimension.

Watch

Extended reading notes

Core claim

RoadFed's central claim is that road-hazard detection can be made accurate, fast, and private at the same time by moving detection to edge servers, federating only quantized model differences to the cloud, and perturbing user data before it leaves devices. The detector MRHD pre-trains separate text and image encoders with a triplet loss that pulls same-class embeddings together and pushes different-class embeddings apart across modalities, then fine-tunes a fused classifier; the authors report this pre-training alone is worth roughly 2-6% accuracy over no pre-training. The federated scheme MFed transmits only int8-quantized weight differences with a decaying learning rate, converging in under 10 global rounds on non-i.i.d. data in their experiments. The privacy scheme MLDP projects high-dimensional inputs into lower-dimensional subspaces with random plus-or-minus-1/e matrices and a Tanh nonlinearity before adding Laplace noise, claiming epsilon-LDP at epsilon=0.8 with acceptable accuracy loss. The paper's headline comparison is that RoadFed reaches 96.42% accuracy and 96.61% F1 at 0.0351 seconds latency with 0.004 GB communication cost, versus 0.29-13.21 GB for the compared systems.

Load-bearing premise

The privacy guarantee rests on the unproven assumption that compressing data through the random matrices before adding noise does not enlarge the sensitivity of the output; if that sensitivity is larger than the Laplace scale assumes, the epsilon-LDP claim for the deployed pipeline collapses.

Editorial extensions

If this is right

  • Hazard alerts can be delivered from edge servers in tens of milliseconds, fast enough for in-range drivers to react before reaching the hazard.
  • Image and text modalities can be fused without requiring paired image-text samples, thanks to triplet relationships across the two encoders.
  • Federated retraining becomes economical on bandwidth-limited links, since only quantized weight differences are sent, about 0.004 GB in the evaluated setting.
  • Strong non-i.i.d. skew is the limiting operating condition: RoadFed's accuracy drops to roughly 20% when each client holds only one or two classes, and recovers above 80% once each client has three or more classes.
  • Privacy at epsilon=0.8 costs roughly 5-12% accuracy relative to noisier settings, while still meeting the real-time latency target.

Reading between the lines

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

  • Beyond the paper's experiments, the dimension-reduction-plus-Laplace recipe, if given a formal sensitivity bound, would extend to other high-dimensional federated sensing data, such as audio or multi-camera feeds, where local differential privacy currently adds prohibitive noise.
  • The reported communication advantage is measured against non-quantized baselines; a direct ablation replacing MFed's quantized-difference step with a quantized FedAvg would isolate how much of the roughly 1,000 times saving comes from the new scheme rather than from quantization itself.
  • The sharp accuracy collapse under extreme non-i.i.d. data suggests a concrete extension: add a small shared public data pool or personalized per-edge layers so that edges holding only one or two classes do not drag the global model to around 20% accuracy.
  • The system's privacy posture assumes untrusted edges and the cloud; if MLDP's guarantee is weakened anywhere, the same architecture would need secure aggregation or encrypted transmission to maintain its stated threat model.
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 / 6 minor

Summary. The paper presents RoadFed, a device-edge-cloud federated learning system for road hazard detection and alarm. It combines three methodological contributions: the Multimodal Road Hazard Detector (MRHD), which fuses image and text features and is trained with a triplet loss; the Multimodal Federated Learning scheme (MFed), which uses adaptive learning-rate decay and dynamic quantization to reduce communication cost; and the Multimodal Local Differential Privacy algorithm (MLDP), which projects high-dimensional inputs to lower dimensions before adding Laplace noise. The authors evaluate the components on a self-collected road hazard dataset, CrisisMMD, and MNIST, and report that the full system achieves 96.42% accuracy with 0.0351 s latency and communication cost as low as 0.004 GB, outperforming several road-detection and federated-learning baselines.

Significance. RoadFed addresses a real application need (timely road hazard alarms) and integrates multimodal learning, federated aggregation, and local privacy in one pipeline. The evaluation is grounded in external and self-collected datasets, uses standard metrics such as accuracy, F1, latency, and communication cost, and compares against a reasonable set of baselines including EcRD, FedRD, FedAvg, FedPAQ, and LRDecay. If the privacy guarantee and the algorithmic details were correct, the communication-efficiency and accuracy results would be of clear value to the C-ITS community. The manuscript also gives detailed experimental parameters and reports variance, which aids reproducibility. The main weakness is that the formal privacy analysis does not cover the algorithm as implemented; this is a load-bearing gap rather than a cosmetic issue.

major comments (3)
  1. [Section 4.3, Eqs. (8)-(13) and Algorithm 2] The claimed epsilon-LDP guarantee is not proved for the mechanism actually run. The proof in Eqs. (10)-(13) is the standard Laplace-mechanism argument for a query f with L1 sensitivity s1(f), and Eq. (8) instantiates f as the identity on the raw d-dimensional input with scale 2d/epsilon. Algorithm 2, however, first maps the data as y' = Tanh(Qy) and x' = Tanh(QxR), with Q and R having entries +/-1/e, and only then adds Laplace noise. The L1 sensitivity of these projected queries is never bounded, so the released y'' and x'' are not shown to satisfy epsilon-LDP. The loop bound j = 1,...,d also does not match the projected dimensions c and c x e. Because privacy preservation is advertised as a central contribution, the authors must either prove a sensitivity bound for the projected maps and calibrate the noise accordingly, or modify Algorithm 2 so that the proof applies.
  2. [Section 5.4, privacy-budget allocation] The per-dimension privacy-budget allocation is inconsistent with the mechanism in Eq. (8). The text states that for 1D text of dimension d, epsilon/d is allocated per dimension, and for 2D images epsilon/d^2 per pixel; yet Eq. (8) adds Laplace noise with scale s1(f)/epsilon to every coordinate of the projected vector, which is the standard construction for epsilon-LDP of the whole vector when s1(f) is the global L1 sensitivity. These two accounting rules cannot both be correct: one treats each coordinate as a separate query with its own budget, while the other treats the full vector as a single query. Please specify the exact query to which epsilon applies and align the noise scale in Algorithm 2 with that choice.
  3. [Algorithm 1, MFed local update] The printed local update, omega_t^i <- -omega_{t-1}^i - (gamma_0/(R+1)) * grad_l(omega_{t-1}^i, b_{t-1}^i), contains a leading minus sign on the previous weights. If taken literally, each local step negates the model, which would prevent the convergence reported in Section 5.3. Please correct the sign (and any related subscript/superscript issues) and confirm that the pseudocode matches the implementation used for the experiments.
minor comments (6)
  1. [Sections 5.3 and 5.5] The references to 'Fig. 5, Fig. 5' are ambiguous and likely refer to different subfigures; renumber or label the subfigures explicitly.
  2. [Section 5.5] 'EdgeRD [9]' is not defined; the citation [9] corresponds to Saha et al., which is listed separately later in the same sentence.
  3. [Section 4.2.2] The phrase 'at time y' is undefined; the round index used elsewhere is t or R.
  4. [Eq. (7)] The roles of zeta and nu are described confusingly ('step size and the last global round'); clarify which variable is the current round and which is the step size.
  5. [Algorithm 2] The loop writes x''[j] even though x' is a c x e array; use element-wise indexing over the projected dimensions.
  6. [Section 5.5] The sentence containing 'it's communication cost' should use the possessive 'its communication cost'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: RoadFed's headline accuracy, latency, and communication cost are measured against independent baselines, and its MLDP proof gap is a correctness issue rather than a circular step.

full rationale

RoadFed's central results are empirical. MRHD accuracy (Tables 4–5) is compared with MobileNetV2, BERT, and multimodal baselines [60–63, 23] on CrisisMMD and a self-collected dataset; MFed convergence and communication cost (Fig. 5, Table 6) are compared with FedAvg, FedPAQ, LRDecay, and earlier road-detection systems; the headline "96.42%" with "0.0351 seconds" and "0.004 GB" communication cost is read directly from Table 6, not produced by fitting a parameter to a desired answer. The privacy proof in Section 4.3 (Eqs. 8–13) reproduces the standard Laplace-mechanism argument from Dwork and Roth [11] for an abstract query f with sensitivity s1(f), and Algorithm 2 then applies Laplace noise to projected quantities such as y' = Tanh(Qy) and x' = Tanh(QxR). The paper never bounds the L1 sensitivity of these projected queries, and the parenthetical in Eq. (8) — "Laplace(s1(f)/ε) means a Laplace distribution with scale 2d/ε" — is valid only for the identity query, not for the Tanh(Q·) projection used in Algorithm 2. This is an omitted proof / unsupported privacy guarantee, and therefore a correctness risk, but it is not circular: the ε-LDP claim is not assumed as its own conclusion, and no fitted value is renamed as a prediction. Self-citations to the authors' earlier works EcRD [4] and FedRD [5] appear as experimental baselines, not as load-bearing justification for RoadFed's derivation. No self-definitional, fitted-input-called-prediction, self-citation-load-bearing, uniqueness-imported, ansatz-smuggled-via-citation, or renaming-known-result circularity is present.

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

The paper introduces no new physical entities, forces, or dimensions; MRHD, MFed, and MLDP are algorithms, not invented entities. The free parameters are mostly hyperparameters chosen by validation or by inspecting test performance.

free parameters (5)
  • triplet loss penalty alpha = 0.1
    Chosen after sweeping alpha in Fig. 4 to maximize accuracy; it weights the text-only triplet loss term in Eq. (3).
  • triplet loss margins c and m = c=0.2, m=0
    Set constant across all experiments.
  • initial learning rate and decay schedule = gamma_0=0.01 / 0.001, delta=0.5, zeta=1
    Used in Eq. (7) for LR decay; values differ per dataset and are chosen by the authors.
  • privacy budget epsilon = 0.8
    Selected from Fig. 7 as the tradeoff between accuracy and privacy after inspecting test accuracy.
  • random projection output dimensions c and e = not fully specified (image downscaled below 1x64)
    Output sizes of Q and R in Algorithm 2 are not given; they determine the dimension reduction and thus the privacy-utility tradeoff.
assumptions (6)
  • standard math The Laplace mechanism provides epsilon-LDP (Dwork-Roth [11])
    Invoked in Section 4.3 to claim that adding Laplace noise to projected data is private.
  • domain assumption Random projection preserves sufficient discriminative information for classification
    Based on Achlioptas [56], used to justify dimension reduction before noise.
  • domain assumption QSGD quantization converges and preserves model quality
    MFed relies on QSGD [36] convergence guarantees; not re-derived.
  • domain assumption Decaying the learning rate improves FedAvg convergence on non-i.i.d. data
    Borrowed from [52] and used to set Eq. (7).
  • domain assumption Edge and cloud servers are untrusted adversaries that may leak or misuse data
    Stated in Section 3.2 and used to motivate MLDP.
  • domain assumption Users in real deployments provide both an image and a text description for each hazard report
    Stated in Section 5.5 to justify multimodal fusion; if false, the system falls back to freezing one modality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RoadFed: A Multimodal Federated Learning System for Improving Road Safety." pith.science (2026). https://pith.science/paper/IGIFXOLD

@misc{pith2026250209978,
  author       = {Pith},
  title        = {Pith review of: RoadFed: A Multimodal Federated Learning System for Improving Road Safety},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IGIFXOLD}},
  note         = {Machine review of arXiv:2502.09978}
}
read the original abstract

Internet of Things (IoTs) have been widely applied in Collaborative Intelligent Transportation Systems (C-ITS) for the prevention of road accidents. As one of the primary causes of road accidents in C-ITS, the efficient detection and early alarm of road hazards are of paramount importance. Given the importance, extensive research has explored this topic and obtained favorable results. However, most existing solutions only explore single-modality data, struggle with high computation and communication overhead, or suffer from the curse of high dimensionality in their privacy-preserving methodologies. To overcome these obstacles, in this paper, we introduce RoadFed, an innovative and private multimodal Federated learning-based system tailored for intelligent Road hazard detection and alarm. This framework encompasses an innovative Multimodal Road Hazard Detector, a communication-efficient federated learning approach, and a customized low-error-rate local differential privacy method crafted for high dimensional multimodal data. Experimental results reveal that the proposed RoadFed surpasses most existing systems in the self-gathered real-world and CrisisMMD public datasets. In particular, RoadFed achieves an accuracy of 96.42% with a mere 0.0351 seconds of latency and its communication cost is up to 1,000 times lower than existing systems in this field. It facilitates collaborative training with non-iid high dimensional multimodal real-world data across various data modalities on multiple edges while ensuring privacy preservation for road users.

Figures

Figures reproduced from arXiv: 2502.09978 by the authors.

Figure 1
Figure 1. An overview of the proposed RoadFed framework, including three key components (i.e., road users’ devices, untrusted [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The proposed Multimodal Road Hazard Detector utilizes a triplet loss to improve feature quality, i.e., enlarging inter-class [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Example images and texts of road hazards (from left [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: MRHD’s accuracy over different α. 5.2. MRHD Results and Evaluation [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Performance comparison of MFed on both the road danger dataset and the MNIST public dataset (MFed-Q and MFed￾LRD refer to MFed without using model quantization and learning rate decay mechanism, respectively) [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Loss of MFed under different local training epochs on our dataset using the MRHD model (each client has 4 classes) [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: The effect of ϵ in MLDP. ϵ = 0.8 is selected as a trade￾off between road hazard detection accuracy and data privacy preservation. 5.5. RoadFed Framework Results and Evaluation RoadFed is evaluated against EcRD [4], EdgeRD [9], Fe￾dRD [5], Vondikakis et al. [26], Dwived…
Figure 10
Figure 10. Figure 10: Detection Result Display, where the red and green markers refer to road areas with and without road hazards. One can [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 60 canonical work pages

  1. [1]

    Z.-S. Tan, E. W. See-To, K.-Y . Lee, H.-N. Dai, M.-L. Wong, Privacy-preserving federated learning for proactive maintenance of iot-empowered multi-location smart city facilities, Journal of Network and Computer Applications 231 (2024) 103996

    work page 2024

  2. [2]

    M. Bakirci, Advanced aerial monitoring and vehicle classification for intelligent transportation systems with yolov8 variants, Journal of Network and Computer Ap- plications (2025) 104134

    work page 2025

  3. [3]

    W. H. Organization, Global status report on road safety 2023, World Health Organization, 2023

    work page 2023

  4. [4]

    Y . Yuan, M. S. Islam, Y . Yuan, S. Wang, T. Baker, L. M. Kolbe, Ecrd: Edge-cloud computing framework for smart road damage detection and warning, IEEE Internet of Things Journal 8 (16) (2020) 12734–12747

    work page 2020

  5. [5]

    Y . Yuan, Y . Yuan, T. Baker, L. M. Kolbe, D. Hogrefe, Fedrd: Privacy-preserving adaptive federated learning framework for intelligent hazardous road damage detec- tion and warning, Future Generation Computer Systems 125 (2021) 385–398

    work page 2021

  6. [6]

    McMahan, E

    B. McMahan, E. Moore, D. Ramage, S. Hampson, B. A. y Arcas, Communication-efficient learning of deep net- works from decentralized data, in: Artificial Intelligence and Statistics, PMLR, 2017, pp. 1273–1282

    2017

  7. [7]

    Zhang, H

    Y . Zhang, H. Zhang, Y . Yang, W. Sun, H. Zhang, Y . Fu, Adaptive differential privacy in asynchronous federated learning for aerial-aided edge computing, Journal of Net- work and Computer Applications 235 (2025) 104087

    work page 2025

  8. [8]

    H. Zhao, Q. Liu, H. Sun, L. Xu, W. Zhang, Y . Zhao, F.- Y . Wang, Community awareness personalized federated learning for defect detection, IEEE Transactions on Com- putational Social Systems (2024)

    work page 2024

Show all 66 references
  1. [9]

    P. K. Saha, D. Arya, Y . Sekimoto, Federated learning– based global road damage detection, Computer-Aided Civil and Infrastructure Engineering 39 (14) (2024) 2223– 2238

    2024

  2. [10]

    Geiping, H

    J. Geiping, H. Bauermeister, H. Dröge, M. Moeller, In- verting gradients-how easy is it to break privacy in feder- ated learning?, Advances in neural information processing systems 33 (2020) 16937–16947

    2020

  3. [11]

    Dwork, A

    C. Dwork, A. Roth, et al., The algorithmic foundations of differential privacy, Foundations and Trends®in Theoret- ical Computer Science 9 (3–4) (2014) 211–407

    2014

  4. [12]

    M. Hao, H. Li, X. Luo, G. Xu, H. Yang, S. Liu, Efficient and privacy-enhanced federated learning for industrial ar- tificial intelligence, IEEE Transactions on Industrial In- formatics 16 (10) (2019) 6532–6542

    2019

  5. [13]

    Xiong, Z

    Z. Xiong, Z. Cai, D. Takabi, W. Li, Privacy threat and defense for federated learning with non-iid data in aiot, IEEE Transactions on Industrial Informatics 18 (2) (2021) 1310–1321

    2021

  6. [14]

    Maeda, Y

    H. Maeda, Y . Sekimoto, T. Seto, T. Kashiyama, H. Omata, Road damage detection and classification using deep neu- ral networks with smartphone images, Computer-Aided Civil and Infrastructure Engineering 33 (12) (2018) 1127– 1141

    2018

  7. [15]

    Moroto, K

    Y . Moroto, K. Maeda, T. Ogawa, M. Haseyama, Snow-or ice-covered road detection in winter road surface condi- tions using deep neural networks, Computer-Aided Civil and Infrastructure Engineering 39 (19) (2024) 2935–2950

    2024

  8. [16]

    Orhan, P

    F. Orhan, P. E. Eren, Road hazard detection and shar- ing with multimodal sensor analysis on smartphones, in: 2013 Seventh International Conference on Next Gener- ation Mobile Apps, Services and Technologies, IEEE, 2013, pp. 56–61

    2013

  9. [17]

    Z.-z. Lan, L. Bao, S.-I. Yu, W. Liu, A. G. Hauptmann, Multimedia classification and event detection using dou- ble fusion, Multimedia tools and applications 71 (2014) 333–347

    2014

  10. [18]

    Kelly, X

    S. Kelly, X. Zhang, K. Ahmad, Mining multimodal infor- mation on social media for increased situational aware- ness, system (2016)

    2016

  11. [19]

    Mouzannar, Y

    H. Mouzannar, Y . Rizk, M. Awad, Damage identification in social media posts using multimodal deep learning., in: ISCRAM, 2018

    2018

  12. [20]

    Abbariki, M

    M. Abbariki, M. Shoman, Interpreting the unexpected: A multimodal framework for out-of-label hazard detection and explanation in autonomous driving, in: Proceedings of the Winter Conference on Applications of Computer Vision, 2025, pp. 669–676

    2025

  13. [21]

    Saeed, M

    N. Saeed, M. Alam, R. G. Nyberg, A multimodal deep learning approach for gravel road condition evaluation through image and audio integration, Transportation En- gineering 16 (2024) 100228

    2024

  14. [22]

    Zhouxiang, O

    L. Zhouxiang, O. Petrosian, Driver assistance system based on multimodal data hazard detection, arXiv preprint arXiv:2502.03005 (2025)

    2025 arXiv

  15. [23]

    Abavisani, L

    M. Abavisani, L. Wu, S. Hu, J. Tetreault, A. Jaimes, Mul- timodal categorization of crisis events in social media, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 14679–14689

    2020

  16. [24]

    X. Dang, X. Shang, Z. Hao, L. Su, Collaborative road damage classification and recognition based on edge com- puting, Electronics 11 (20) (2022) 3304

    2022

  17. [25]

    Y . Liu, F. Liu, Y . Huang, J. Hu, W. Zhang, Y . Hou, The real-time pavement distress detection system based on edge-cloud collaborative computing, IEEE Transactions on Intelligent Transportation Systems (2025)

    2025

  18. [26]

    I. V . V ondikakis, I. E. Panagiotopoulos, G. J. Dimi- trakopoulos, Fedrsc: A federated learning analysis for multi-label road surface classifications, IEEE Open Jour- nal of Intelligent Transportation Systems (2024)

    2024

  19. [27]

    H.-T. Wu, H. Li, H.-L. Chi, W.-B. Kou, Y .-C. Wu, S. Wang, A hierarchical federated learning framework for collaborative quality defect inspection in construction, Engineering Applications of Artificial Intelligence 133 (2024) 108218

    2024

  20. [28]

    S. K. Dwivedi, D. Arya, Y . Sekimoto, Road damage de- tection across borders: Federated learning insights from japan, china, norway and the usa, in: 2024 IEEE Smart World Congress (SWC), IEEE, 2024, pp. 1446–1452

    2024

  21. [29]

    Pauly, D

    L. Pauly, D. Hogg, R. Fuentes, H. Peel, Deeper networks for pavement crack detection, in: Proceedings of the 34th ISARC, IAARC, 2017, pp. 479–485

    2017

  22. [30]

    K. Ma, M. Hoai, D. Samaras, Large-scale continual road inspection: Visual infrastructure assessment in the wild., in: BMVC, 2017

    2017

  23. [31]

    R. Wang, F. He, W. Yang, L. Zhao, Assistant driving safety early warning system based on internet of vehicles, in: International Conference on Machine Learning and Big Data Analytics for IoT Security and Privacy, Springer, 2021, pp. 969–976

    2021

  24. [32]

    Sulistyowati, A

    R. Sulistyowati, A. Suryowinoto, H. Sujono, I. Iswahyudi, Monitoring of road damage detection systems using im- age processing methods and google map, in: IOP Con- ference Series: Materials Science and Engineering, V ol. 1010, IOP Publishing, 2021, p. 012017

    2021

  25. [33]

    Maeda, N

    K. Maeda, N. Ogawa, T. Ogawa, M. Haseyama, Damage- level classification considering both correlation between image and text data and confidence of attention map, Computer-Aided Civil and Infrastructure Engineering 40 (6) (2025) 764–781

    2025

  26. [34]

    H. Tian, Y . Feng, M. Quddus, Y . Demiris, P. Angeloudis, Multimodal learning for traffic risk prediction: Combin- ing aerial imagery with contextual data, IEEE Open Jour- nal of Intelligent Transportation Systems (2025)

    2025

  27. [35]

    H. B. McMahan, F. Yu, P. Richtarik, A. Suresh, D. Bacon, et al., Federated learning: Strategies for improving com- munication efficiency, in: Proceedings of the 29th Confer- ence on Neural Information Processing Systems (NIPS), Barcelona, Spain, 2016, pp. 5–10

    2016

  28. [36]

    Alistarh, D

    D. Alistarh, D. Grubic, J. Li, R. Tomioka, M. V ojnovic, Qsgd: Communication-efficient sgd via gradient quantiza- tion and encoding, Advances in Neural Information Pro- cessing Systems 30 (2017) 1709–1720

    2017

  29. [37]

    Bernstein, Y .-X

    J. Bernstein, Y .-X. Wang, K. Azizzadenesheli, A. Anand- kumar, signsgd: Compressed optimisation for non-convex problems, in: International Conference on Machine Learning, PMLR, 2018, pp. 560–569

    2018

  30. [38]

    Y . Zhao, M. Li, L. Lai, N. Suda, D. Civin, V . Chan- dra, Federated learning with non-iid data, arXiv preprint arXiv:1806.00582 (2018)

    2018 arXiv

  31. [39]

    X. Li, H. Zhao, W. Deng, Iofl: Intelligent-optimization- based federated learning for non-iid data, IEEE Internet of Things Journal 11 (9) (2024) 16693–16699

    2024

  32. [40]

    Z. Li, Y . Sun, J. Shao, Y . Mao, J. H. Wang, J. Zhang, Fea- ture matching data synthesis for non-iid federated learn- ing, IEEE Transactions on Mobile Computing 23 (10) (2024) 9352–9367

    2024

  33. [41]

    Kairouz, K

    P. Kairouz, K. Bonawitz, D. Ramage, Discrete distribution estimation under local privacy, in: International Confer- ence on Machine Learning, PMLR, 2016, pp. 2436–2444

    2016

  34. [42]

    T. Wang, J. Blocki, N. Li, S. Jha, Locally differen- tially private protocols for frequency estimation, in: 26th {USENIX}Security Symposium ({USENIX}Security 17), 2017, pp. 729–745

    2017

  35. [43]

    G. C. Fanti, V . Pihur, Ú. Erlingsson, Building a rappor with the unknown: Privacy-preserving learning of associ- ations and data dictionaries., Proc. Priv. Enhancing Tech- nol. 2016 (3) (2016) 41–61

    2016

  36. [44]

    Ren, C.-M

    X. Ren, C.-M. Yu, W. Yu, S. Yang, X. Yang, J. A. Mc- Cann, S. Y . Philip, Lopub: High-dimensional crowd- sourced data publication with local differential privacy, IEEE Transactions on Information Forensics and Security 13 (9) (2018) 2151–2166

    2018

  37. [45]

    C. Xia, J. Hua, W. Tong, S. Zhong, Distributed k-means clustering guaranteeing local differential privacy, Com- puters & Security 90 (2020) 101699

    2020

  38. [46]

    Batool, A

    H. Batool, A. Anjum, A. Khan, S. Izzo, C. Mazzocca, G. Jeon, A secure and privacy preserved infrastructure for vanets based on federated learning with local differential privacy, Information Sciences 652 (2024) 119717

    2024

  39. [47]

    Y .-z. Li, L. Xu, J. Zhang, et al., Wf-ldpsr: A local differ- ential privacy mechanism based on water-filling for secure release of trajectory statistics data, Computers & Security 148 (2025) 104165

    2025

  40. [48]

    Hussain, W

    A. Hussain, W. Akbar, T. Hussain, A. K. Bashir, M. M. Al Dabel, F. Ali, B. Yang, Ensuring zero trust iot data privacy: Differential privacy in blockchain using feder- ated learning, IEEE Transactions on Consumer Electron- ics (2024)

    2024

  41. [49]

    Joachims, T

    T. Joachims, T. Finley, C.-N. J. Yu, Cutting-plane training of structural svms, Machine learning 77 (1) (2009) 27–59

    2009

  42. [50]

    B. Shaw, B. Huang, T. Jebara, Learning a distance metric from a network, Advances in Neural Information Process- ing Systems 24 (2011) 1899–1907

    2011

  43. [51]

    Reisizadeh, A

    A. Reisizadeh, A. Mokhtari, H. Hassani, A. Jadbabaie, R. Pedarsani, Fedpaq: A communication-efficient feder- ated learning method with periodic averaging and quanti- zation, in: International Conference on Artificial Intelli- gence and Statistics, PMLR, 2020, pp. 2021–2031

    2020

  44. [52]

    X. Li, K. Huang, W. Yang, S. Wang, Z. Zhang, On the convergence of fedavg on non-iid data, arXiv preprint arXiv:1907.02189 (2019)

    2019 arXiv

  45. [53]

    Gholami, S

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, K. Keutzer, A survey of quantization methods for efficient neural network inference, in: Low-Power Computer Vi- sion, Chapman and Hall/CRC, 2022, pp. 291–326

    2022

  46. [54]

    Pytorch, The documents of dynamic quantization in py- torch,https://pytorch.org/tutorials/recipes/ recipes/dynamic_quantization.html(2021 (last accessed January 11, 2024))

    2021

  47. [55]

    Kairouz, S

    P. Kairouz, S. Oh, P. Viswanath, Extremal mechanisms for local differential privacy, in: Advances in neural informa- tion processing systems, 2014, pp. 2879–2887

    2014

  48. [56]

    Achlioptas, Database-friendly random projections, in: Proceedings of the twentieth ACM SIGMOD-SIGACT- SIGART symposium on Principles of database systems, 2001, pp

    D. Achlioptas, Database-friendly random projections, in: Proceedings of the twentieth ACM SIGMOD-SIGACT- SIGART symposium on Principles of database systems, 2001, pp. 274–281

    2001

  49. [57]

    F. Alam, F. Ofli, M. Imran, Crisismmd: Multimodal twit- ter datasets from natural disasters, in: Proceedings of the international AAAI conference on web and social media, V ol. 12, 2018

    2018

  50. [58]

    Sandler, A

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L.-C. Chen, Mobilenetv2: Inverted residuals and linear bottle- necks, in: Proceedings of the IEEE conference on com- puter vision and pattern recognition, 2018, pp. 4510– 4520

    2018

  51. [59]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre- training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  52. [60]

    F. Ofli, F. Alam, M. Imran, Analysis of social media data using multimodal deep learning for disaster response, arXiv preprint arXiv:2004.11838 (2020)

    2020 arXiv

  53. [61]

    Gallo, A

    I. Gallo, A. Calefati, S. Nawaz, M. K. Janjua, Image and encoded text fusion for multi-modal classification, in: 2018 Digital Image Computing: Techniques and Applica- tions (DICTA), IEEE, 2018, pp. 1–7

    2018

  54. [62]

    Choi, J.-S

    J.-H. Choi, J.-S. Lee, Embracenet for activity: A deep multimodal fusion architecture for activity recognition, in: Adjunct Proceedings of the 2019 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2019 ACM International Symposium ...

    2019

  55. [63]

    R. Pranesh, Exploring multimodal features and fusion strategies for analyzing disaster tweets, in: Proceedings of the Eighth Workshop on Noisy User-generated Text (W- NUT 2022), 2022, pp. 62–68

    2022

  56. [64]

    Dwork, F

    C. Dwork, F. McSherry, K. Nissim, A. Smith, Calibrating noise to sensitivity in private data analysis, in: Theory of cryptography conference, Springer, 2006, pp. 265–284

    2006

  57. [65]

    Erlingsson, V

    Ú. Erlingsson, V . Pihur, A. Korolova, Rappor: Random- ized aggregatable privacy-preserving ordinal response, in: Proceedings of the 2014 ACM SIGSAC conference on computer and communications security, 2014, pp. 1054– 1067

    2014

  58. [66]

    T. T. Nguyên, X. Xiao, Y . Yang, S. C. Hui, H. Shin, J. Shin, Collecting and analyzing data from smart de- vice users with local differential privacy, arXiv preprint arXiv:1606.05053 (2016)

    2016 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.