Pith. sign in

REVIEW 5 major objections 7 minor 32 references

SecureV2X: An Efficient and Privacy-Preserving System for Vehicle-to-Everything (V2X) Applications

T0 review · 5 major / 7 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 SecureV2X claims private V2X inference can run near real time, up to 100x faster than prior secure systems.

desk verdict A genuinely useful MPC-engineering contribution for V2X, but the ReLU/ELU mismatch and baseline-comparison problems keep the headline claims from being trustworthy as written. read the letter →

arxiv 2508.19115 v1 pith:MZ6PW3LQ submitted 2025-08-26 cs.CR cs.AI

Joshua Lee , Ali Arastehfard , Weiran Liu , Xuegang Ban , Yuan Hong This is my paper
classification cs.CRcs.AI
keywords PrivacyDrowsinessDetectionRed-lightViolationVehicle-to-Everything(V2X)SecureMulti-PartyComputationSecretSharingCrypTenYOLOv5
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

SecureV2X claims that secure multi-party computation can make two safety-critical V2X applications practical in real time: drowsiness detection from single-channel EEG and red-light violation detection from traffic-camera video. The paper builds a two-party secret-sharing system, CryptoDrowsy and FastSec-YOLO, where a server holds proprietary model weights and the vehicle or camera holds private data, and neither side learns the other's input. The system adds missing secure operations such as upsampling, split, and constant floor, plus an extension to the CrypTen graph processor, then reports dramatic speedups over prior secure versions of these models. If the claims hold, privacy-preserving V2X inference becomes a viable candidate for deployment rather than a theoretical cost burden.

What carries the argument

The key mechanism is a two-party secret-sharing inference pipeline: all values are additive shares over a ring, Beaver triples (⟨a⟩,⟨b⟩,⟨c⟩ with c=ab) enable secure multiplication, and comparisons and activations are computed by converting to binary shares (A2B/B2A) and back. CrypTen provides the primitive library; the paper's contribution is adding three missing secure functionalities (upsampling, tensor split, constant floor) and an extension to the graph processor that accepts empty-argument ONNX functions, which together let YOLOv5 and CompactCNN run in the secure domain. These custom operations are mostly non-interactive, which keeps the round count low.

What would settle it

Run the plaintext CompactCNN as defined in Section 2.1 (with ELU) and the secure protocol exactly as listed in Algorithm 1 (which invokes ReLU) on the same EEG test set; if the secure output's accuracy does not match the reported 0.8184, the parity claim is unsupported for the protocol as written.

Watch

Extended reading notes

Core claim

The paper's central claim is that SecureV2X is the first real-time privacy-preserving system for EEG-based driver drowsiness detection and YOLOv5-based red-light violation detection. Using additive secret sharing, Beaver triples from a mediating agent, and arithmetic/binary share conversions, CryptoDrowsy and FastSec-YOLO let a client and server run inference so the client learns only the output and the server learns nothing about the input. The authors report accuracy parity with plaintext for drowsiness (0.8184) and precision comparable to a plaintext YOLOv3 pipeline for red-light detection. Compared to prior secure systems, CryptoDrowsy is 9.4x faster, uses 143x fewer rounds, and 16.6x le

Load-bearing premise

The claimed accuracy parity holds only if the secure protocol computes the exact CompactCNN model that achieves the reported plaintext accuracy; the protocol in Algorithm 1 substitutes ReLU where the architecture in Section 2.1 specifies ELU, and the paper never explains that substitution.

Editorial extensions

If this is right

  • CryptoDrowsy completes a 314-sample EEG batch in about 0.027 seconds, so private driver monitoring becomes plausible on edge servers.
  • FastSec-YOLO completes a 32-image batch in about 1.7 seconds on a GPU, over 100x faster than prior secure object detectors, moving red-light enforcement closer to real-time.
  • The multi-agent evaluation shows runtime grows only marginally from 1 to 15 simultaneous clients, indicating the mediating agent can serve many vehicles and cameras at once.
  • Because upsampling, split, and constant floor are now secure, the framework can be extended to other network architectures that need these tensor operations.

Reading between the lines

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

  • The paper leaves implicit that part of the 100x speedup over prior secure object detectors comes from choosing YOLOv5n, the smallest YOLOv5 variant; its own Table 1 shows larger variants give better accuracy but several times slower runtime, so a like-for-like comparison against Faster-RCNN/YOLOv3 with matched accuracy would narrow the gap.
  • The reported runtimes do not appear to include the mediating agent's cost of generating Beaver triples; a full deployment would need to add that preprocessing, which is a testable extension of the reported numbers.
  • The secure upsampling and split operations would also apply to segmentation and other dense-prediction networks, so the protocol suite is likely reusable beyond the two demonstrated V2X tasks.
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

5 major / 7 minor

Summary. This paper presents SecureV2X, a secure two-party computation system built on CrypTen for two V2X applications: EEG-based driver drowsiness detection using the CompactCNN model, and red-light violation detection using YOLOv5n. The authors design two protocols, CryptoDrowsy and FastSec-YOLO, add secure implementations of operations missing from CrypTen (upsampling, split, constant floor), and evaluate accuracy, runtime, communication, and multi-agent scaling. The headline claims are that CryptoDrowsy is 9.4x faster, uses 143x fewer rounds, and involves 16.6x less communication than prior secure drowsiness systems, and that FastSec-YOLO is roughly 100x faster than prior secure object detectors.

Significance. If the efficiency and accuracy claims survive scrutiny, SecureV2X would be a meaningful engineering contribution: it demonstrates that a compact EEG CNN and YOLOv5n can be instantiated under secret sharing, and that multiple clients can be served simultaneously with modest overhead. The open-source release, use of a standard framework (CrypTen), and evaluation of mixed multi-agent workloads are clear strengths. However, the central claims currently rest on an activation-function inconsistency, an uninterpretable round count, and cross-paper baseline comparisons that are not demonstrated to be apples-to-apples. The contribution is potentially publishable after major revisions.

major comments (5)
  1. [§2.1, Algorithm 1, §3.1] Section 2.1 specifies CompactCNN with ELU activation (equation after Eq. (1)), and the surrounding text states 'The ELU activation function is applied over this output.' However, Algorithm 1 line 4 calls F_ReLU, Section 3.1 says 'ReLU activation employs a multiplexer', and Section 5 lists ReLU among implemented functionalities. No justification is given for replacing ELU with ReLU. Since ELU and ReLU differ on negative inputs, the secure protocol may be evaluating a different classifier from the plaintext CompactCNN. Consequently, the reported accuracy parity in Section 6.1 (0.8184 for CrypTen, CrypTFlow2, and plaintext) does not establish that the private system preserves the accuracy of the stated model. Please either implement ELU securely or clearly justify the substitution and report accuracy for the actual secure model.
  2. [Table 2, §6.1] Table 2 reports CryptoDrowsy's round complexity as 0.411. A protocol's round complexity is normally an integer count of communication rounds per execution; 0.411 suggests the value was amortized over the batch of 314 EEG signals. If so, the 143x rounds reduction claimed in the abstract compares an amortized per-input figure to CrypTFlowDrowsy's 59 rounds, whose basis (per input, per batch, or per full inference) is not stated. Please report un-amortized round counts for a single inference for all systems, and clarify the basis for every round count. Without this, the '143x fewer computational rounds' claim is not interpretable.
  3. [Table 3, §6.2] Table 3 compares FastSec-YOLO's 1.734 s runtime to 180 s, 190.527 s, and 276 s for SecRCNN, P2OD, and PPDF. The footnote says prior works' reported results are used because code is unavailable. These numbers come from different papers, presumably with different hardware, input resolutions, batch sizes, and software stacks; no environment or normalization is reported for them. The 'nearly 100x faster' claim in the abstract and Section 6.2 is therefore not yet substantiated as a head-to-head comparison. Please provide full hardware/software details for all runs, and either re-run the baselines where possible or clearly label the comparison as indicative and state the original conditions.
  4. [§5.2] Section 5.2 states: 'Each party secretly transmits its arithmetic share to the other and then performs addition over both shares.' If each of the two parties sends its arithmetic share to the other, both parties end up with both shares and can reconstruct the plaintext; this is not a secure A2B conversion and contradicts the security claims of Section 5. Please rewrite this paragraph to describe the actual CrypTen conversion protocol (e.g., local conversion of each arithmetic share to a binary secret share followed by a secure adder), or provide a correct citation. As written, this part of the security argument is invalid.
  5. [§5, Contributions] The contributions list claims 'provable security' and Section 5 argues that security follows from composition of CrypTen primitives, citing [11]. However, no security model, formal functionality definitions, or theorem statement is given for CryptoDrowsy or FastSec-YOLO. A generic appeal to composition is insufficient, especially given the A2B issue above. Please add a precise statement of the adversary model and a modular security argument, or temper the 'provable security' claim to 'security inherited from CrypTen under the semi-honest model'.
minor comments (7)
  1. [§4.1] 'Additional details about its construction can be seen in figure 2' refers to Figure 2, which is an experimental plot, not a protocol diagram. Please fix the cross-reference.
  2. [Algorithm 2] Line 23 says 'P1 sends its share <O_n>_k to P0', but the subscript should be the party index (e.g., <O_n>_1), not the generic k. As written, it is ambiguous which share is sent.
  3. [Table 1] The caption lists 'Mean Precision (MP)' as a column, but the table body has no MP column. Add the column or remove the caption entry.
  4. [Algorithm 7] Line 4 has a missing closing parenthesis: 'F_MaxPool2D(<t2>k' should be 'F_MaxPool2D(<t2>k)'.
  5. [§1] Duplicated word: 'This This drastically increases...' should be 'This drastically increases...'.
  6. [§6.1] The sentence 'CrypTen and CrypTFlow2 achieve the same results in plaintext at 0.8394' is unclear—do you mean the F1-score? Please reword.
  7. [§6] The experimental environment is not described: GPU model, CPU, memory, and software versions are missing. This is important for reproducibility and for interpreting the runtime comparisons.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; SecureV2X's claims rest on external models, external MPC infrastructure, and direct benchmarks. The ELU/ReLU discrepancy is a consistency concern, not a circular reduction.

full rationale

I find no load-bearing circular step. The derivation chain is: adopt CompactCNN (Cui et al.) and YOLOv5 (Yahya/Ultralytics) as plaintext models; implement private inference using CrypTen primitives; benchmark against external or independently constructed baselines (Delphi, CrypTFlow2, SecRCNN, P2OD, PPDF, CrypTen-YOLOv5 variants). None of these inputs are defined in terms of the paper's outputs, and no parameter is fitted so that a reported speedup or accuracy is forced by construction. The accuracy and efficiency numbers are experimental measurements, not quantities obtained by renaming the inputs. The one substantive internal inconsistency is that Section 2.1 specifies CompactCNN's activation as ELU (O_{i+1} = ELU(Batch(Conv(...)))), while Algorithm 1 line 4 invokes F_ReLU and Section 3.1 describes 'ReLU activation employs a multiplexer'. This suggests the secure protocol may not implement the exact same network as the 0.8184 plaintext CompactCNN, so the parity claim is not fully supported. However, that is a protocol-consistency/correctness gap, not a circularity: the secure result is not equivalent to the plaintext result by definition, and no equation-level self-reference is present. There is also no load-bearing self-citation chain: the cited MPC infrastructure (CrypTen, ABY, GMW, etc.) is external to this paper, and no 'uniqueness' or 'first' claim is justified by the authors' own prior theorems. The paper is self-contained against external benchmarks, so I score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on a semi-honest trusted-helper security model, a composition-security assumption, representativeness of small benchmark datasets, and architectural fidelity of the secure models. The evaluation hyperparameters are hand-picked. No new physical entities are invented.

free parameters (3)
  • input image size = 288x288
    Chosen as optimal trade-off between speed and accuracy for FastSec-YOLO based on the same experiments used to report results (Section 6.2).
  • batch size = 32 (Tables 1 and 3), 16 described as optimal
    Batch size affects inference time; 32 is used for comparisons and selected via empirical analysis in Section 6.2.
  • EEG evaluation batch = 314 signals
    All drowsiness results use a batch of 314 EEG signals; no dataset split or cross-validation is described.
assumptions (4)
  • domain assumption Semi-honest two-party model with a trusted mediating agent that distributes Beaver triples
    Both protocols assume parties follow the protocol and that the helper is available; security claims are conditional on this, stated in Sections 4.1 and 5.
  • standard math Composition security theorem of [11] applies to the customized CrypTen primitives used here
    Section 5 derives system security from composition of secure primitives without verifying the theorem's preconditions for upsampling, split, floor, and A2B/B2A conversions.
  • domain assumption COCO128 and the 314-sample EEG batch are representative of real red-light and drowsiness deployment conditions
    Accuracy and timing claims are based on these datasets; no deployment-specific or multi-site benchmark is used (Section 6).
  • domain assumption The secure model architecture matches the plaintext CompactCNN and YOLOv5n architectures
    Accuracy parity assumes identical architecture, but the ELU/ReLU mismatch between Section 2.1 and Algorithm 1 indicates this may not hold.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SecureV2X: An Efficient and Privacy-Preserving System for Vehicle-to-Everything (V2X) Applications." pith.science (2026). https://pith.science/paper/MZ6PW3LQ

@misc{pith2026250819115,
  author       = {Pith},
  title        = {Pith review of: SecureV2X: An Efficient and Privacy-Preserving System for Vehicle-to-Everything (V2X) Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MZ6PW3LQ}},
  note         = {Machine review of arXiv:2508.19115}
}
abstract

Autonomous driving and V2X technologies have developed rapidly in the past decade, leading to improved safety and efficiency in modern transportation. These systems interact with extensive networks of vehicles, roadside infrastructure, and cloud resources to support their machine learning capabilities. However, the widespread use of machine learning in V2X systems raises issues over the privacy of the data involved. This is particularly concerning for smart-transit and driver safety applications which can implicitly reveal user locations or explicitly disclose medical data such as EEG signals. To resolve these issues, we propose SecureV2X, a scalable, multi-agent system for secure neural network inferences deployed between the server and each vehicle. Under this setting, we study two multi-agent V2X applications: secure drowsiness detection, and secure red-light violation detection. Our system achieves strong performance relative to baselines, and scales efficiently to support a large number of secure computation interactions simultaneously. For instance, SecureV2X is $9.4 \times$ faster, requires $143\times$ fewer computational rounds, and involves $16.6\times$ less communication on drowsiness detection compared to other secure systems. Moreover, it achieves a runtime nearly $100\times$ faster than state-of-the-art benchmarks in object detection tasks for red light violation detection.

Figures

Figures reproduced from arXiv: 2508.19115 by the authors.

Figure 1
Figure 1. Overview for the SecureV2X Framework. Server: edge server or the cloud (infrastructure), Client: connected vehicle [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Evaluation of FastSec-YOLO on Varying Batch and Image Sizes under the Secure and Plaintext Settings. We compute [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Evaluation of SecureV2X on Simultaneously Supporting Multiple Vehicles (in different ratios). Each subfigure displays [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 29 canonical work pages

  1. [11]

    Ivan Damgård, Valerio Pastro, Nigel Smart, and Sarah Zakarias. 2012. Multiparty computation from somewhat homomorphic encryption. In Proceedings of the 32nd Annual Cryptology Conference on Advances in Cryptology . Springer Berlin Heidelberg, 643–662

    work page 2012

  2. [1]

    AAA. 2019. Red light running crash fatalities. (2019). https://us.vocuspr.com /Newsroom/ViewAttachment.aspx?SiteName=AAACS&Entity=PRAsset&At tachmentType=F&EntityID=110661&AttachmentID=a05c7218-b5a8-4f9e-86 8d-557e5959f030

    work page 2019

  3. [2]

    McKinney, Dongrui Wu, Chin- Teng Lin, Martine De Cock, and Anderson C

    Anisha Agarwal, Rafael Dowsley, Nicholas D. McKinney, Dongrui Wu, Chin- Teng Lin, Martine De Cock, and Anderson C. A. Nascimento. 2019. Protecting Privacy of Users in Brain-Computer Interface Applications. IEEE Transactions on Neural Systems and Rehabilitation Engineering , 27, 8, (Aug. 2019), 1546–1555. Conference Name: IEEE Transactions on Neural System...

    work page arXiv 2019

  4. [3]

    Khaled R. Ahmed. 2021. Smart Pothole Detection Using Deep Learning Based on Dilated Convolution. Sensors, 21

    work page 2021

  5. [4]

    Donald Beaver. 1991. Efficient multiparty protocols using circuit randomization. In 11th Annual International Cryptology Conference . Springer, 420–432

    work page 1991

  6. [5]

    Renwan Bi, Jinbo Xiong, Youliang Tian, Qi Li, and Kim-Kwang Raymond Choo

    work page

  7. [6]

    Dan Bogdanov, Sven Laur, and Jan Willemson. 2008. Sharemind: a framework for fast privacy-preserving computations. In Computer Security - ESORICS 2008 . Springer Berlin Heidelberg, 192–206

    work page 2008

  8. [7]

    Cohn, Suman Kakar, Chloe Perkins, Rebecca Steinbach, and Phil Edwards

    Ellen G. Cohn, Suman Kakar, Chloe Perkins, Rebecca Steinbach, and Phil Edwards. 2020. Red light camera interventions for reducing traffic violations and traffic crashes: a systematic review. Campbell Systematic Reviews , 16, 2, 1–52

    work page 2020

Show all 32 references
  1. [8]

    Ronald Cramer, Ivan Damgård, and Yuval Ishai. 2005. Share conversion, pseu- dorandom secret-sharing and applications to secure computation. In Theory of Cryptography: Second Theory of Cryptography Conference, TCC 2005, Cambridge, MA, USA, February 10-12, 2005. Proceedings 2 . ...

    2005

  2. [9]

    Jian Cui, Zirui Lan, Yisi Liu, Ruilin Li, Fan Li, Olga Sourina, and Wolfgang Müller-Wittig. 2022. A compact and interpretable convolutional neural network for cross-subject driver drowsiness detection from single-channel EEG. en. Methods, 202, (June 2022), 173–184. doi: 10.101...

    2022 doi

  3. [10]

    Ivan Damgård, Fitzi Matthias, Eike Kiltz, Jesper Buus Nielsen, and Tomas Toft. 2006. Unconditionally secure constant-rounds multi-party computation for equality, comparison, bits and exponentiation. In Proceedings of the Third Conference on Theory of Cryptography . Springer Be...

    2006

  4. [12]

    Daniel Demmler, Thomas Schneider, and Michael Zohner. 2015. Aby - a frame- work for efficient mixed-protocol secure two-party computation. In NDSS Symposium 2015. NDSS, 1–15

    2015

  5. [13]

    Oded Goldreich, Silvio Micali, and Avi Widerson. 1987. How to play any men- tal game, or a completeness theorem for protocols with honest majority. In Proceedings of the nineteenth annual ACM symposium on Theory of computing . STOC, 218–229

    1987

  6. [14]

    Zdenek Kalal, Krystian Mikolajczyk, and Jiri Matas. 2010. Forward-Backward Error: Automatic Detection of Tracking Failures. In 2010 20th International Conference on Pattern Recognition . IEEE, 2756–2759

    2010

  7. [15]

    Aicha Khalfaoui, Abdelmajid Badri, and Ilham EL Mourabit. 2022. Comparative Study of YOLOv3 and YOLOv5’s Performances for Real-time Person Detection. In 2022 2nd International Conference on Innovative Research in Applied Science, Engineering and Technology (IRASET). IEEE, 1–5

    2022

  8. [16]

    Brian Knott, Shobha Venkataraman, Awni Hannun, Shubho Sengupta, Mark Ibrahim, and Laurens van der Maaten. 2022. CrypTen: Secure Multi-Party Computation Meets Machine Learning. en. arXiv:2109.00984 [cs]. (Sept. 2022). Retrieved Aug. 5, 2024 from http://arxiv.org/abs/2109.00984

    2022 arXiv

  9. [17]

    Chin-Teng Lin, Ruei-Cheng Wu, Sheng-Fu Liang, Wen-Hung Chao, Yu-Jie Chen, and Tzyy-Ping Jung. 2005. Eeg-based drowsiness estimation for safety driving using independent component analysis. IEEE Transactions on Circuits and Systems 1, 52, 12, 2726–2738

    2005

  10. [18]

    Tsung-Yi Lin et al. 2015. Microsoft coco: common objects in context. (2015). arXiv: 1405.0312 [cs.CV]

    2015 arXiv

  11. [19]

    Yang Liu, Zhuo Ma, Ximeng Liu, Siqi Ma, and Kui Ren. 2022. Privacy-Preserving Object Detection for Medical Images With Faster R-CNN. IEEE Transactions on Information Forensics and Security , 17, 69–84. Conference Name: IEEE Transac- tions on Information Forensics and Security....

    2022

  12. [20]

    Upesh Nepal and Hossein Eslamiat. 2022. Comparing YOLOv3, YOLOv4, and YOLOv5 for Autonomous Landing Spot Detection in Faulty UAVs.Sensors, 22, 2, 1–15

    2022

  13. [21]

    US Department of Transportation. 2024. Usdot releases national deployment plan for vehicle-to-everything (v2x) technologies to reduce death and serious injuries on america’s roadways. (2024). https://www.transportation.gov/briefi ng-room/usdot-releases-national-deployment-plan...

    2024

  14. [22]

    Rishabh Poddar, Ganesh Ananthanarayanan, Srinath Setty, Stavros Volos, and Raluca Ada Popa. 2020. Visor: Privacy-Preserving video analytics as a cloud service. In 29th USENIX Security Symposium (USENIX Security 20) . USENIX Association, (Aug. 2020), 1039–1056. isbn: 978-1-9391...

    2020

  15. [23]

    Andreea Bianca Popescu, Ioana Antonia Taca, Cosmin Ioan Nita, Anamaria Vizitiu, Robert Demeter, Constantin Suciu, and Lucian Mihai Itu. 2021. Privacy Preserving Classification of EEG Data Using Machine Learning and Homo- morphic Encryption. en. Applied Sciences, 11, 16, (Jan. ...

    2021 doi

  16. [24]

    Muhammad Ramzan, Hilkmat Ullah Khan, Shahid Mahmood Awan, Amina Ismail, Mahwish Ilyas, and Mahmood Ahsan. 2019. A survey on state-of-the-art drowsiness detection techniques. IEEE Access, 7, 61904–61919

    2019

  17. [25]

    Shah, Dhaval K

    Foram N. Shah, Dhaval K. Patel, Kashish D. Shah, Mehul S. Raval, Mukesh Zaveri, and S.N. Merchant. 2023. Novel Crash Prevention Framework for C-V2X using Deep Learning. In 2023 15th International Conference on COMmunication Systems & NETworkS (COMSNETS) . ISSN: 2155-2509. (Jan...

    2023

  18. [26]

    David M Studdert, Simon J Walter, and Jeremy J Goldhaber-Fiebert. 2017. Once ticketed, twice shy? specific deterrence from road traffic laws. In Health Law Workshops. Harvard Law School

    2017

  19. [27]

    Le Quang Thao, Duong Duc Cuong, Nguyen Tuan Anh, Pham Mai Anh, Ha Minh Duc, and Nguyen Minh. 2022. Automatic Traffic Red-Light Violation Detection Using AI. en. Ingénierie des systèmes d information , 27, 1, (Feb. 2022), 75–80. doi: 10.18280/isi.270109

    2022 doi

  20. [28]

    Ahmad Yahya, Ihab Abdelkareem, and Osama Al-fakhouri. [n. d.] Fully-automated- red-light-violation-detection. (). https://github.com/AhmadYahya97/Fully-Aut omated-red-light-Violation-Detection/tree/master

  21. [29]

    Muhamad Munawarar Yusro, Rozniza Ali, and Muhammad Suzuri Hitam. 2023. Comparison of Faster R-CNN and YOLOv5 for Overlapping Objects Recogni- tion. Baghdad Science Journal, 20, 3, 893–903

    2023

  22. [30]

    Rui Zhang, Cong Xie, and Liwei Deng. 2023. A fine-grained object detection model for aerial images based on yolov5 deep neural network. Chinese Journal of Electronics, 32, 1, 51–63

    2023

  23. [31]

    Yongjie Zhou, Jinbo Xiong, Renwan Bi, and Youliang Tian. 2022. Secure YOLOv3- SPP; Edge-Cooperative Privacy-preserving Object Detection for Connected Autonomous Vehicles. In 2022 International Conference on Networking and Applications. IEEE, (Dec. 2022), 82–89. Conference’17, ...

    2022

  24. [2023]

    IEEE Internet of Things Journal , 10, 3, (Feb

    Achieving Lightweight and Privacy-Preserving Object Detection for Connected Autonomous Vehicles. IEEE Internet of Things Journal , 10, 3, (Feb. 2023), 2314–2329. Conference Name: IEEE Internet of Things Journal. doi: 10.1109/JIOT.2022.3212464

    2023

Pith tools

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