The DOI matches its Crossref record; 79 of the 82 references it cites were found in a registry or an index; 11 of the 35 sources its claims cite were checked for support.
Event Anonymization: Privacy-Preserving Person Re-Identification and Pose Estimation in Event-Based Vision
Open accessAbstract
This work presents an end-to-end network architecture jointly optimized for the twofold objective of preserving privacy and performing a downstream computer vision task.
Overview
CoChat Summary
## TL;DR
The paper proposes an end-to-end event-stream anonymization framework for event cameras that aims to block image-reconstruction privacy attacks while still supporting downstream tasks. It is evaluated on person re-identification and human pose estimation, plus several privacy attacks, and reports that anonymization preserves utility with only small performance drops while greatly degrading reconstructed-image quality.
## Key findings
- On the proposed Event-ReId dataset, the anonymized model reaches Rank-1 = 72.7% and mAP = 47.5%, compared with 78.2% / 52.7% without privacy.
- Compared with event encryption baselines on Event-ReId, the proposed method has much lower reconstructed-image scores: SSIM 0.376 and PSNR 8.795, versus 0.471/10.18 for “Encryption Discarding” and 0.436/10.06 for “Encryption Scrambling.”
- For privacy attacks on Event-ReId, face verification AUC drops to 0.517 under privacy, versus 0.748 for RGB and 0.699 for non-private reconstructed images.
- For HPE on DHP19, the paper reports small drops: 2D joint error differences of 0.32 pixels (camera 2) and 0.27 pixels (camera 3), and 3D MPJPE changes from 67.29 mm to 72.53 mm.
- Under gender prediction on Event-ReId real data, privacy reduces gender accuracy to 26.7% (16/60) and detection accuracy to 33.3% (20/60); on SoftBio synthetic data, privacy gives 18.4% detection and 6.6% gender accuracy.
## Methods
- The method uses a three-part pipeline: an event anonymization block (convolutional autoencoder), an image reconstruction block (pre-trained E2VID as the attacker), and a downstream task block for person ReId and HPE.
- Inputs are event voxel grids; the ReId backbone is a modified ResNet-50 producing 256-D embeddings and trained with cross-entropy + triplet loss, while HPE uses heatmap prediction plus soft-argmax and is trained with JSD + geometric loss.
- Evaluation uses Event-ReId-v2 (60 identities, 4 cameras), SoftBio synthetic ReId data, and DHP19 for HPE; the paper also tests image retrieval, face recognition, gender prediction, inversion attack, and adversarial learning.
## Limitations
- The paper states that the method is designed to preserve downstream performance, but it still reports some drops: for example, ReId Rank-1 falls from 78.2% to 72.7%, and HPE 3D MPJPE increases from 67.29 mm to 72.53 mm.
- It explicitly notes a limitation of the comparison to prior work for person ReId: “there is no other method for direct comparison” for privacy-preserving person ReId beyond the discussed baselines.
- Evident from the text: the evaluation is limited to the datasets and tasks reported here (Event-ReId/SoftBio/DHP19), so generalization beyond these settings is not established in the provided text.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
9 Valid · 3 Not valid · 18 Not checked · 46 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“Du et al. [12] recently introduced a manually designed encryption framework to counteract privacy attacks on event data. This method uses spatial chaotic mapping to scramble event positions and invert their polarities.”
What the sources say
- Supportedcites [12] source
Section 4 describes the authors’ proposed encryption framework: a two-dimensional chaotic map generates sequences used to scramble event positions and flip polarities. Section 5.1 evaluates the method against visualization attacks on event data. The indexed passages provide section labels but no page numbers. (LLM, conf=0.93)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Our proposed dataset demonstrates a favorable size relative to existing event-based datasets, notably in the domain of activity recognition n-HAR [16] and DailyAction-DVS [17], and human pose estimation dataset DHP19 [2] (see comparisons in Table 1).”
What the sources say
- Not checked — no full text of the sourcecites [16] source
- Supportedcites [17] source
The citation identifies DailyAction-DVS as an existing event-based action-recognition dataset. Section 2.3 states that this paper proposes DailyAction-DVS and that it contains 1,440 recordings. The favorable-size comparison is the citing authors’ own conclusion, not a finding attributed to this source. (LLM, conf=0.91)
- Not checked — no full text of the sourcecites [2] source
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“As illustrated in Figure 1(a), event-toimage [8] model reconstruct a detailed image that reveals personal identity information; consequently, this implies that event cameras can no longer be deemed inherently privacypreserving.”
What the sources say
- Partly supportedcites [8] sourceExperimental — accuracy not yet measured
The Introduction (p. 2) says the model reconstructs high-quality images from event data, and §5.2 (p. 8) reports reconstructing an entire scene in a selfie sequence. Our copy does not establish that the reconstructed image reveals personal identity information; that is an additional inference, not a finding the paper reports. (LLM, conf=0.82)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“As illustrated in Figure 1(a), event-toimage [8] model reconstruct a detailed image that reveals personal identity information; consequently, this implies that event cameras can no longer be deemed inherently privacypreserving.”
What the sources say
- Partly supportedcites [8] sourceExperimental — accuracy not yet measured
The Evaluation (p. 6) reports detailed image reconstructions, and Section 5.2 (p. 8) describes reconstructing an entire scene from a selfie sequence. The paper does not establish that the images reveal personal identity information or assess whether event cameras are privacy-preserving; those are the citing authors’ inferences. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Our proposed event anonymization method outperforms the event encryption technique.”
What the sources say
- Supportedcites [12] source
Section 4.2 describes the paper’s event encryption algorithm, so it supports the citation’s identification of an event encryption technique. The statement that the citing authors’ anonymization method outperforms it is their own comparison, not a result attributed to this source. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
35 more claims · 3 with their cited sources checked
Create a free account to see all 76 claims and to verify the 3 references still open.
- Probably not a substantial claim. 6
- It may state two separate claims as one. 30
- The quote only partly states this claim. 2
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 2×
“We utilize the same backbone we use for person ReId with the same settings, except we remove ResNet-50 [59] layers after the second residual block following [3].”
MethodArcFace: Additive Angular Margin Loss for Deep Face Recognition
Deng, J., Guo, J., Xue, N., & Zafeiriou, S. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00482 · Cited by 7,781 · cited 1×
“We use the pre-trained face recognition model ArcFace [65] to measure the resilience of our system to this privacy attack.”
MethodOpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields
Cao, Z., Hidalgo, G., Simon, T., Wei, S.-E., & Sheikh, Y. (2021) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2019.2929257 · Cited by 5,536
Image Quality Metrics: PSNR vs. SSIM
Hore, A., & Ziou, D. (2010) · 2010 20th International Conference on Pattern Recognition · doi:10.1109/icpr.2010.579 · Cited by 4,943 · cited 2×
“SSIM [58], [62] is a metric used to measure the similarity between two images.”
MethodIn Defense of the Triplet Loss for Person Re-Identification
Hermans, A., Beyer, L., & Leibe, B. (2017) · arXiv (Cornell University) · doi:10.48550/arxiv.1703.07737 · Cited by 2,687 · cited 2×
“In addition, triplet loss is one of the most popular depths metric losses in person ReId [63] that further improves the discriminative property by increasing the inter-class discrepancy while decreasing intra-class distinctness : where D ap is the distance between the anchor and the positive sample…”
BackgroundMethodGallego, G., Delbruck, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A. J., Conradt, J., Daniilidis, K., et al. (2022) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2020.3008413 · Cited by 2,566
Deep Learning for Person Re-Identification: A Survey and Outlook
Ye, M., Shen, J., Lin, G., Xiang, T., Shao, L., & Hoi, S. C. H. (2022) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2021.3054775 · Cited by 2,259
Bag of Tricks and a Strong Baseline for Deep Person Re-Identification
Luo, H., Gu, Y., Liao, X., Lai, S., & Jiang, W. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw.2019.00190 · Cited by 1,584 · cited 2×
“For this experiment, we employ the state-of-the-art person ReId model BOT [64] to evaluate image retrieval and use the test sets of real and simulated datasets.”
MethodA Low Power, Fully Event-Based Gesture Recognition System
Amir, A., Taba, B., Berg, D., Melano, T., McKinstry, J., Di Nolfo, C., Nayak, T., Andreopoulos, A., Garreau, G., Mendoza, M., et al. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.781 · Cited by 1,017 · cited 1×
“Among these, human pose estimation [2] dataset, action recognition [16], [17], [18] dataset, face expression recognition [19] dataset, and car recognition [20], [21] datasets.”
BackgroundConverting Static Image Datasets to Spiking Neuromorphic Datasets Using Saccades
Orchard, G., Jayawant, A., Cohen, G. K., & Thakor, N. (2015) · Frontiers in Neuroscience · doi:10.3389/fnins.2015.00437 · Cited by 882 · cited 2×
“Semi-synthetic datasets such as those in [23] and [24] convert standard video into event data.”
BackgroundMethod
Cited by
Event Camera-Based Eye Motion Analysis: A Survey
Iddrisu, K., Shariff, W., Corcoran, P., O'Connor, N. E., Lemley, J., & Little, S. (2024) · IEEE Access · doi:10.1109/access.2024.3462109 · Cited by 34
“Another study proposed an event scrambling technique that makes the stream uninterpretable to the human eye while still allowing effective application of computer vision models [36].”
BackgroundGuo, Y., Gao, T., Dong, A., Jiang, X., Zhu, Z., & Wang, F. (2025) · Sensors · doi:10.3390/s25082409 · Cited by 44
“To address this, researchers have explored privacy-preserving techniques, such as on-device processing [140], anonymization of human pose data [141,142], and differential privacy methods [143], ensuring that surveillance systems operate within ethical boundaries while maintaining their…”
A review of recent techniques for person re-identification
Asperti, A., Fiorilla, S., Nardi, S., & Orsini, L. (2024) · Machine Vision and Applications · doi:10.1007/s00138-024-01622-3 · Cited by 31
“The potentiality and limitations of this approach are discussed in [128].”
A Review on Research and Application of AI-Based Image Analysis in the Field of Computer Vision
Wu, P., He, X., Dai, W., Zhou, J., Shang, Y., Fan, Y., & Hu, T. (2025) · IEEE Access · doi:10.1109/access.2025.3565300 · Cited by 34
“Reference [93] enhances pedestrian trajectory prediction efficiency in dynamic environments by employing region-based clustering for group motion analysis.”
Recoverable Anonymization for Pose Estimation: A Privacy-Enhancing Approach
Huang, W., Ni, Y., Rezvani, A., Jeong, S., Chen, H., Liu, Y., Wen, F., & Imani, M. (2025) · IEEE Winter Conference on Applications of Computer Vision · doi:10.1109/wacv61041.2025.00512 · Cited by 13
“For privacy enhancement in computer vision applications, a straightforward solution is to use very low-resolution data [39, 54].”
Metadata
- Authors
- Ahmad, S., Morerio, P., & Del Bue, A.
- Affiliations
- Italian Institute of Technology (IT)
- Pattern Analysis and Computer Vision (PAVIS), Istituto Italiano di Tecnologia at Italy (IIT), Genoa, Italy
- Venue
- IEEE Access
- Published
- 2024
- Volume
- 12
- Pages
- 66964-66980
- ISSN
- 2169-3536 (ISSN-L)
- OpenAlex
- W4396817472
- Semantic Scholar
- 97f722bc9db48c0573021ba559b6d8ea595dd1ae
- Cited by
- 16
- References
- 82
- Citations by year
- 2026: 7 · 2025: 7 · 2024: 2
- FWCI
- 2.25
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- IEEE Access (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Surveillance and Tracking Methods
- Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Vision and Imaging
- Keywords
- event cameras, privacy preservation, person re-identification, human pose estimation, visual surveillance, neuromorphic vision
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Event Anonymization: Privacy-Preserving Person Re-Identification and Pose Estimation in Event-Based Vision. https://app.cochat.ai/papers/10.1109/access.2024.3399539
- The paper
- Ahmad, S., Morerio, P., & Del Bue, A. (2024). Event Anonymization: Privacy-Preserving Person Re-Identification and Pose Estimation in Event-Based Vision. IEEE Access. https://doi.org/10.1109/access.2024.3399539
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.