pith. machine review for the scientific record. sign in

arxiv: 2604.25544 · v1 · submitted 2026-04-28 · 💻 cs.CR · cs.AI

Recognition: unknown

Medoid Prototype Alignment for Cross-Plant Unknown Attack Detection in Industrial Control Systems

Luyao Wang
Authors on Pith no claims yet

Pith reviewed 2026-05-07 15:38 UTC · model grok-4.3

classification 💻 cs.CR cs.AI
keywords medoid prototype alignmentcross-plant detectionunknown attack detectionindustrial control systemsdomain shiftintrusion detectiontransfer learningICS security
0
0 comments X

The pith

Medoid prototype alignment lets an ICS intrusion detector trained on one plant detect unknown attacks on another by compressing traffic and matching stable summaries.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a framework that maps site-specific industrial control system traffic into a shared representation space, then extracts medoid prototypes to capture typical operational patterns in source and target domains. A prototype-calibrated transfer objective aligns the target medoids to source medoids while keeping source labels discriminative and encouraging confident predictions on the target. This avoids noisy direct sample matching that fails under heterogeneous ICS conditions. A sympathetic reader cares because real deployments face scarce labels for new attacks and traffic that differs sharply between plants, so a method that transfers without retraining each site would cut deployment cost and delay. If correct, the approach delivers the highest average accuracy and F1-score across four unknown-attack transfer tasks between natural-gas and water-storage systems.

Core claim

By first compressing heterogeneous ICS traffic into a comparable representation space and then extracting robust medoid prototypes that summarize local operational structure, the method aligns target prototypes to source prototypes via a calibrated transfer objective; this preserves source-domain discrimination, reduces noisy cross-domain matching, and yields stable detection of unseen attacks, reaching an average accuracy of 0.843 and average F1-score of 0.838 on four cross-plant transfer tasks while revealing clear transfer asymmetry that favors prototype guidance in harder reverse directions.

What carries the argument

The medoid prototype alignment framework, which extracts medoid prototypes to summarize each domain's local structure and applies a prototype-calibrated transfer objective to align them without direct sample matching.

If this is right

  • The method reduces noisy cross-domain matching and improves transfer stability under heterogeneous industrial conditions.
  • Prototype guidance proves especially helpful on challenging reverse-transfer settings where source and target roles are swapped.
  • The approach achieves the best average performance across compared models on the four unknown-attack tasks between gas and water control systems.
  • Transfer asymmetry between directions is observable and can be mitigated by the prototype mechanism.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same prototype-summarization step could be applied to other domain-shift detection problems where full sample alignment is too noisy, such as cross-organization network monitoring.
  • If medoids prove stable, the framework might allow incremental addition of new plants without full retraining, only periodic prototype updates.
  • The observed asymmetry suggests testing whether the method still works when the target plant has fewer normal-operation samples than the source.

Load-bearing premise

Heterogeneous ICS traffic from different plants can be compressed into one representation space where medoid prototypes reliably capture operational structure and remain stable enough for cross-domain alignment without erasing source discrimination.

What would settle it

Running the method on a new pair of ICS plants with substantially different protocol mixes or attack distributions and finding that accuracy falls below the best baseline or that medoid alignment produces more false positives than direct feature alignment.

Figures

Figures reproduced from arXiv: 2604.25544 by Luyao Wang.

Figure 1
Figure 1. Figure 1: Teaser of the proposed medoid prototype alignment idea. The figure summarizes the cross-plant view at source ↗
Figure 2
Figure 2. Figure 2: Prototype extraction and cross-domain matching in the updated medoid prototype alignment module. For each target prototype 𝑞 𝑡 ℓ , a soft correspondence over source prototypes is computed: 𝑎ℓ𝑘 = exp(−∥𝑞 𝑡 ℓ − 𝑞 𝑠 𝑘 ∥ 2 2 /𝜏) Í 𝑢 exp(−∥𝑞 𝑡 ℓ − 𝑞 𝑠 𝑢∥ 2 2 /𝜏) . The prototype alignment loss is then defined as L𝑝𝑟𝑜𝑡𝑜 = ∑︁ 𝐾𝑡 ℓ=1 ∑︁ 𝐾𝑠 𝑘=1 𝑎ℓ𝑘 ∥𝑞 𝑡 ℓ − 𝑞 𝑠 𝑘 ∥ 2 2 . To preserve source discriminability, we use s… view at source ↗
Figure 3
Figure 3. Figure 3: Updated overall pipeline of the proposed medoid prototype alignment framework for cross-plant view at source ↗
Figure 4
Figure 4. Figure 4: Average cross-task performance with standard-deviation error bars. MPA achieves the best mean behavior on both Accuracy and F1-score view at source ↗
Figure 5
Figure 5. Figure 5: Worst-case robustness derived from the four transfer tasks. Left: minimum Accuracy and F1-score achieved by each model. Right: across-task performance range, where smaller values indicate more consistent behavior. should achieve both a high worst-case score and a compact range. MPA achieves the strongest worst-case behavior, maintaining at least 0.81 Accuracy and 0.80 F1-score even on its hardest task. At … view at source ↗
Figure 7
Figure 7. Figure 7: Task-level comparison between MPA and baseline models on four unknown-attack transfer tasks view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative illustration of domain alignment view at source ↗
read the original abstract

Deploying an intrusion detector trained in one industrial plant to another remains difficult because Industrial Control System (ICS) traffic is highly site-dependent, labels are scarce, and unseen attacks often appear after deployment. To address this challenge, this paper introduces a medoid prototype alignment framework for cross-plant unknown attack detection. Instead of aligning all source and target samples directly, the method first compresses heterogeneous traffic into a comparable representation space and then extracts robust medoid prototypes that summarize local operational structure in each domain. A prototype-calibrated transfer objective is further designed to align target prototypes with source prototypes while preserving source-domain discrimination and encouraging confident target predictions. This strategy reduces noisy cross-domain matching and improves transfer stability under heterogeneous industrial conditions. Experiments conducted on natural gas and water storage control systems show that the proposed method achieves the best average performance among all compared models, reaching an average accuracy of 0.843 and an average F1-score of 0.838 across four unknown-attack transfer tasks. The analysis also shows clear transfer asymmetry between source-target directions and confirms that prototype guidance is especially helpful on challenging reverse-transfer settings. These findings suggest that medoid prototype alignment is a practical solution for robust industrial intrusion detection under domain shift.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a medoid prototype alignment framework for cross-plant unknown attack detection in ICS. Heterogeneous traffic is first compressed into a comparable representation space; robust medoid prototypes are then extracted to summarize local operational structure in each domain. A prototype-calibrated transfer objective aligns target prototypes with source prototypes while preserving source-domain discrimination and encouraging confident target predictions. Experiments on natural gas and water storage control systems report that the method attains the best average performance across four unknown-attack transfer tasks, with mean accuracy 0.843 and mean F1-score 0.838, and note transfer asymmetry favoring certain source-target directions.

Significance. If the empirical superiority holds under rigorous validation, the work supplies a practical prototype-based strategy for domain adaptation in ICS intrusion detection, where site-dependent traffic and post-deployment unseen attacks are common. The emphasis on medoid summarization rather than direct sample alignment could improve stability in heterogeneous settings; the reported results on two real control-system datasets provide a concrete starting point for further benchmarking.

major comments (2)
  1. [Abstract] Abstract: the headline performance claim (average accuracy 0.843, F1 0.838 across four unknown-attack transfer tasks) is presented without error bars, statistical significance tests, ablation results, or even a high-level description of the four tasks and the baseline models, rendering the superiority assertion impossible to assess for post-hoc tuning or missing controls.
  2. [Method] Method description (abstract and §3): the claim that medoid prototypes 'summarize local operational structure' and that the calibrated objective 'reduces noisy cross-domain matching' rests on the unverified assumption that the initial compression step produces a representation space in which site-specific sensor correlations and protocol artifacts do not dominate; no sensitivity analysis, feature-robustness check, or visualization of the learned space is supplied to support this load-bearing premise.
minor comments (2)
  1. [Abstract] Abstract: the datasets are referred to only as 'natural gas and water storage control systems' without naming the public repositories or providing basic statistics (number of flows, attack types, etc.).
  2. [Method] The paper would benefit from a short pseudocode block or explicit equations for the prototype extraction and the calibrated transfer loss.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract's empirical claims and the method's foundational assumptions. We address both points directly below, providing clarifications from the manuscript and indicating revisions to strengthen the presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline performance claim (average accuracy 0.843, F1 0.838 across four unknown-attack transfer tasks) is presented without error bars, statistical significance tests, ablation results, or even a high-level description of the four tasks and the baseline models, rendering the superiority assertion impossible to assess for post-hoc tuning or missing controls.

    Authors: We acknowledge the abstract's brevity limits context. The full manuscript (Sections 4.1 and 4.2) defines the four tasks as bidirectional transfers (gas-to-water and water-to-gas) on two distinct unknown attack types from the gas and water ICS datasets, with baselines including DANN, ADDA, and other domain adaptation methods. Error bars (std. dev. over 5 runs), paired t-tests for significance, and ablation results (Section 5) are reported in the paper. In revision, we will expand the abstract with a one-sentence description of the tasks and baselines plus a reference to the detailed statistical and ablation results in the main text. This improves assessability while respecting length constraints. revision: partial

  2. Referee: [Method] Method description (abstract and §3): the claim that medoid prototypes 'summarize local operational structure' and that the calibrated objective 'reduces noisy cross-domain matching' rests on the unverified assumption that the initial compression step produces a representation space in which site-specific sensor correlations and protocol artifacts do not dominate; no sensitivity analysis, feature-robustness check, or visualization of the learned space is supplied to support this load-bearing premise.

    Authors: Section 3.1 specifies that the shared autoencoder is trained jointly on both domains to learn a latent space that captures invariant operational features from heterogeneous ICS traffic, thereby reducing the impact of site-specific sensor correlations and protocol artifacts before medoid extraction. The prototype alignment objective in §3.3 then operates on these summaries to avoid direct sample matching. To further validate, the revised manuscript will add: t-SNE visualizations of the latent space (showing preserved operational clusters with reduced domain shift), a sensitivity analysis on latent dimension and encoder depth (demonstrating stable medoid quality), and a robustness check via controlled perturbation of sensor features. These additions directly support the premise without altering the original design rationale. revision: yes

Circularity Check

0 steps flagged

No circularity: framework description contains no self-referential reductions or fitted inputs renamed as predictions

full rationale

The provided abstract and description introduce a medoid prototype alignment method by outlining sequential steps (compression into representation space, medoid extraction, prototype-calibrated transfer objective) without any equations, self-citations, or uniqueness theorems that would make the claimed accuracy/F1 gains equivalent to the inputs by construction. No step reduces a 'prediction' to a fitted parameter or renames a known pattern; the performance numbers are presented as experimental outcomes on specific ICS datasets rather than derived tautologies. The central claims rest on empirical comparison rather than load-bearing self-referential logic, satisfying the criteria for a self-contained non-circular presentation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on the domain assumption that traffic from different plants can be mapped to a shared space where medoids capture stable structure; no free parameters or invented entities are named in the abstract.

axioms (2)
  • domain assumption Heterogeneous ICS traffic from different plants can be compressed into a comparable representation space
    Invoked as the first step of the framework in the abstract.
  • domain assumption Medoid prototypes summarize local operational structure robustly enough to support stable alignment
    Central to the claim that prototype alignment reduces noisy matching.

pith-pipeline@v0.9.0 · 5506 in / 1374 out tokens · 36436 ms · 2026-05-07T15:38:15.611191+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 11 canonical work pages · 2 internal anchors

  1. [1]

    Standardize source and target traffic features

  2. [2]

    Apply PCA to obtainZ𝑠 andZ 𝑡

  3. [3]

    Extract medoid prototypesP 𝑠 andP 𝑡 using K-Medoids

  4. [4]

    Encode samples and prototypes with𝑓𝜃

  5. [5]

    Compute prototype correspondences𝑎ℓ 𝑘

  6. [6]

    OptimizeL 𝑠𝑢 𝑝 +𝛼L 𝑝𝑟 𝑜𝑡𝑜 +𝛽L 𝑒𝑛𝑡

  7. [7]

    Four unknown-attack transfer tasks are considered: •DoS(G)→NMRI(W), •SMRI(G)→MPCI(W), •DoS(W)→NMRI(G), •SMRI(W)→MPCI(G)

    Predict target labels with𝑔𝜙 (𝑓 𝜃 (𝑧 𝑡 𝑗)) •RQ2: How sensitive are different models to transfer direction between plants? •RQ3: Does prototype guidance contribute additional robustness beyond standard transfer learning? Experimentsarebuiltontwoindustrialsystems: a naturalgascontrolsystem(G)andawaterstoragetank control system (W). Four unknown-attack trans...

  8. [8]

    A theory of learning from different domains.Machine Learning, 79(1–2):151–175, 2010

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. A theory of learning from different domains.Machine Learning, 79(1–2):151–175, 2010

    2010

  9. [9]

    Organ-agents: Virtual human physiology simulator via llms

    Rihao Chang, Hongbo Jiao, Weizhi Nie, Huijie Guo,KaiXie,ZihanWu,LinZhao,YutongBai, Yongtao Ma, Lijuan Wang, et al. Organ-agents: Virtual human physiology simulator via llms. arXiv preprint arXiv:2508.14357, 2025

    work page arXiv 2025

  10. [10]

    3d shape knowledge graph for cross-domain 3d shape retrieval

    Rihao Chang, Yongtao Ma, Tong Hao, Weijie Wang, and Weizhi Nie. 3d shape knowledge graph for cross-domain 3d shape retrieval. CAAI Transactions on Intelligence Technology, 9(5):1199–1216, 2024

    2024

  11. [11]

    Y. Chen, S. Su, D. Yu, H. He, X. Wang, Y. Ma, and H. Guo. Cross-domain industrial intrusion detection deep model trained with imbalanced data.IEEE Internet of Things Journal, 10:584–596, 2023

    2023

  12. [12]

    An llm agent for automatic geospatial data analysis.arXiv preprint arXiv:2410.18792, 2024

    Yuxing Chen, Weijie Wang, Sylvain Lobry, and Camille Kurtz. An llm agent for automatic geospatial data analysis.arXiv preprint arXiv:2410.18792, 2024

    work page arXiv 2024

  13. [13]

    Domain-adversarial training of neural networks.Journal of Machine Learning Research, 17(59):1–35, 2016

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks.Journal of Machine Learning Research, 17(59):1–35, 2016

    2016

  14. [14]

    Point cloud completionguidedbypriorknowledgeviacausal inference.arXiv preprint arXiv:2305.17770, 2023

    Songxue Gao, Chuanqi Jiao, Ruidong Chen, Weijie Wang, and Weizhi Nie. Point cloud completionguidedbypriorknowledgeviacausal inference.arXiv preprint arXiv:2305.17770, 2023

    work page arXiv 2023

  15. [15]

    Point-pc: Point cloud completion guided by prior knowledge via causal inference.CAAI Transactions on Intelligence Technology, 2025

    Xuesong Gao, Chuanqi Jiao, Ruidong Chen, Weijie Wang, and Weizhi Nie. Point-pc: Point cloud completion guided by prior knowledge via causal inference.CAAI Transactions on Intelligence Technology, 2025

    2025

  16. [16]

    M. R. Gauthama Raman, Chuadhry Mujeeb Ahmed, and Aditya Mathur. Machine learning for intrusion detection in industrial control systems: Challenges and lessons from experimental evaluation.Cybersecurity, 4(1):27, 2021

    2021

  17. [17]

    Jolliffe.Principal Component Analysis

    Ian T. Jolliffe.Principal Component Analysis. Springer, New York, 2 edition, 2002

    2002

  18. [18]

    Rousseeuw

    Leonard Kaufman and Peter J. Rousseeuw. Partitioning around medoids (program pam). In Finding Groups in Data: An Introduction to Cluster Analysis, pages 68–125. John Wiley & Sons, New York, 1990

    1990

  19. [19]

    Deep transfer learning for intrusion detection in industrial control networks: A comprehensive review.Journal of Network and Computer Applications, 220:103760, 2023

    HamzaKheddar,YassineHimeur,andAliIsmail Awad. Deep transfer learning for intrusion detection in industrial control networks: A comprehensive review.Journal of Network and Computer Applications, 220:103760, 2023

    2023

  20. [20]

    Detecting cyberattacks in industrial control systems using convolutional neural networks.arXiv preprint arXiv:1806.08110, 2018

    Moshe Kravchik and Asaf Shabtai. Detecting cyberattacks in industrial control systems using convolutional neural networks.arXiv preprint arXiv:1806.08110, 2018

    work page arXiv 2018

  21. [21]

    Freeinsert: Disentangled text-guided object insertion in 3d gaussian scene without spatial priors, 2025

    Chenxi Li, Weijie Wang, Qiang Li, Bruno Lepri, Nicu Sebe, and Weizhi Nie. Freeinsert: Disentangled text-guided object insertion in 3d gaussian scene without spatial priors, 2025

    2025

  22. [22]

    Multimodal information fusion based on lstm for 3d model retrieval.Multimedia Tools and Applications, 79(45–46):33943–33956, 2020

    Qi Liang, Ning Xu, Weijie Wang, and Xingjian Long. Multimodal information fusion based on lstm for 3d model retrieval.Multimedia Tools and Applications, 79(45–46):33943–33956, 2020

    2020

  23. [23]

    Rfr-wwanet: Weighted window attention-based recovery feature resolution network for unsupervised image registration.Pattern Recognition, 2023

    Mingrui Ma, Tao Wang, Liyuan Song, Weijie Wang, and Guixia Liu. Rfr-wwanet: Weighted window attention-based recovery feature resolution network for unsupervised image registration.Pattern Recognition, 2023

    2023

  24. [24]

    Mingrui Ma, Weijie Wang, Jie Ning, Jianfeng He,NicuSebe,andBrunoLepri.Largelanguage models for multimodal deformable image registration.arXiv preprint arXiv:2408.10703, 2024

    work page arXiv 2024

  25. [25]

    Unsupervised deep probabilistic approach for partial point cloud registration

    Guofeng Mei, Hao Tang, Xiaoshui Huang, Weijie Wang, Juan Liu, Jian Zhang, Luc Van Gool, and Qiang Wu. Unsupervised deep probabilistic approach for partial point cloud registration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  26. [26]

    Haroon Mushtaq, S. U. Khan, M. A. Jan, A. Ullah, and H. A. Khattak. A parallel architectureforthepartitioningaroundmedoids (pam) algorithm.Sensors, 18(12):4129, 2018

    2018

  27. [27]

    T2td: Text-3d generation modelbasedonpriorknowledgeguidance.IEEE TPAMI, 2024

    WeizhiNie,RuidongChen,WeijieWang,Bruno Lepri, and Nicu Sebe. T2td: Text-3d generation modelbasedonpriorknowledgeguidance.IEEE TPAMI, 2024

    2024

  28. [28]

    T2td: Text-3d generation modelbasedonpriorknowledgeguidance.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(1):172–189, 2025

    WeizhiNie,RuidongChen,WeijieWang,Bruno Lepri, and Nicu Sebe. T2td: Text-3d generation modelbasedonpriorknowledgeguidance.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(1):172–189, 2025

    2025

  29. [29]

    Characteristic views extraction modalbased-ondeepreinforcementlearningfor 3d model retrieval

    Weizhi Nie, Weijie Wang, Anan Liu, and Chuang Chen. Characteristic views extraction modalbased-ondeepreinforcementlearningfor 3d model retrieval. In2019 IEEE International Conference on Image Processing (ICIP), pages 2389–2393, 2019

    2019

  30. [30]

    Hgan: Holistic generative adversarial networks for two-dimensional image-based three-dimensional object retrieval

    Weizhi Nie, Weijie Wang, Anan Liu, Jie Nie, and Yuxuan Su. Hgan: Holistic generative adversarial networks for two-dimensional image-based three-dimensional object retrieval. ACM Transactions on Multimedia Computing, Communications,andApplications,15(4):1–24, 2019

    2019

  31. [31]

    Domainadaptationviatransfer component analysis

    SinnoJialinPan,IvorW.Tsang,JamesT.Kwok, andQiangYang. Domainadaptationviatransfer component analysis. InProceedings of the 21stInternationalJointConferenceonArtificial Intelligence (IJCAI), pages 1187–1192, 2009

    2009

  32. [32]

    On lines and planes of closest fit to systems of points in space.The London, Edinburgh,andDublinPhilosophicalMagazine and Journal of Science, 2(11):559–572, 1901

    Karl Pearson. On lines and planes of closest fit to systems of points in space.The London, Edinburgh,andDublinPhilosophicalMagazine and Journal of Science, 2(11):559–572, 1901

    1901

  33. [33]

    Bringing masked autoencoders explicit contrastive properties for point cloud self-supervised learning

    Bin Ren, Guofeng Mei, Danda Pani Paudel, Weijie Wang, Yawei Li, Mengyuan Liu, Rita Cucchiara, Luc Van Gool, and Nicu Sebe. Bringing masked autoencoders explicit contrastive properties for point cloud self-supervised learning. InProceedings of the Asian Conference on Computer Vision (ACCV), 2024

    2024

  34. [34]

    A. Rigo, L. Stornaiuolo, Weijie Wang, M. Martino, Bruno Lepri, Nicu Sebe, and Weizhi Nie. Poci-diff: Position objects consistently and interactively with 3d-layout guided diffusion.arXiv preprint arXiv:2601.14056, 2026

    work page arXiv 2026

  35. [35]

    MuhammadAzmiUmer,KhurumNazirJunejo, Muhammad Taha Jilani, and Aditya P. Mathur. Machine learning for intrusion detection in industrial control systems: Applications, challenges,andrecommendations.International Journal of Critical Infrastructure Protection, 38:100516, 2022

    2022

  36. [36]

    Mbt-polyp: A new multi-branch memory-augmented transformer for polyp segmentation.Image and Vision Computing, 163:105747, 2025

    Tao Wang, Weijie Wang, Fausto Giunchiglia, Fengzhi Zhao, Ye Zhang, Duo Yu, and Guixia Liu. Mbt-polyp: A new multi-branch memory-augmented transformer for polyp segmentation.Image and Vision Computing, 163:105747, 2025

    2025

  37. [37]

    U-hrmlp: Refining segmentation boundaries in histopathology images

    Tao Wang, Kai Zhang, Weijie Wang, Mingrui Ma, Ye Zhang, He Zhao, and Guixia Liu. U-hrmlp: Refining segmentation boundaries in histopathology images. In2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1–5, 2024

    2024

  38. [38]

    Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptivesemanticsegmentation

    WeiWang,ZhunZhong,WeijieWang,XiChen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptivesemanticsegmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24090–24099, 2023

    2023

  39. [39]

    Fully-geometric cross-attention for point cloud registration

    Weijie Wang, Guofeng Mei, Jian Zhang, Nicu Sebe, Bruno Lepri, and Fabio Poiesi. Fully-geometric cross-attention for point cloud registration. In3DV, 2025

    2025

  40. [40]

    Fully-geometric cross-attention for point cloud registration

    Weijie Wang, Guofeng Mei, Jian Zhang, Nicu Sebe, Bruno Lepri, and Fabio Poiesi. Fully-geometric cross-attention for point cloud registration. In3DV. IEEE, 2025

    2025

  41. [41]

    Zeroreg: Zero-shot point cloudregistrationwithfoundationmodels.arXiv preprint arXiv:2312.03032, 2023

    Weijie Wang, Wenqi Ren, Guofeng Mei, Bin Ren, Xiaoshui Huang, Fabio Poiesi, Nicu Sebe, and Bruno Lepri. Zeroreg: Zero-shot point cloudregistrationwithfoundationmodels.arXiv preprint arXiv:2312.03032, 2023

    work page arXiv 2023

  42. [42]

    Rethinking the learning paradigm for facial expression recognition.arXiv preprint arXiv:2209.15402, 2022

    Weijie Wang, Nicu Sebe, and Bruno Lepri. Rethinking the learning paradigm for facial expression recognition.arXiv preprint arXiv:2209.15402, 2022

    work page arXiv 2022

  43. [43]

    PoInit-of-View: Poisoning Initialization of Views Transfers Across Multiple 3D Reconstruction Systems

    Weijie Wang, Songlong Xing, Zhengyu Zhao, Nicu Sebe, and Bruno Lepri. Poinit-of-view: Poisoninginitializationofviewstransfersacross multiple 3d reconstruction systems.arXiv preprint arXiv:2604.16540, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Uvmap-id: A controllable and personalized uv map generative model

    Weijie Wang, Jichao Zhang, Chang Liu, Xia Li, Xingqian Xu, Humphrey Shi, Nicu Sebe, and Bruno Lepri. Uvmap-id: A controllable and personalized uv map generative model. InACM MM, pages 10725–10734, 2024

    2024

  45. [45]

    Turn fake into real: Adversarial head turn attacks against deepfake detection

    Weijie Wang, Zhengyu Zhao, Nicu Sebe, and Bruno Lepri. Turn fake into real: Adversarial head turn attacks against deepfake detection. arXiv preprint arXiv:2309.01104, 2023

    work page arXiv 2023

  46. [46]

    Learning spatial-spectral dual adaptive graph embedding formultispectralandhyperspectralimagefusion

    Xuquan Wang, Feng Zhang, Kai Zhang, Weijie Wang, Xiong Dun, and Jiande Sun. Learning spatial-spectral dual adaptive graph embedding formultispectralandhyperspectralimagefusion. Pattern Recognition, 151:110365, 2024

    2024

  47. [47]

    Finetune Like You Pretrain: Boosting Zero-shot Adversarial Robustness in Vision-language Models

    Songlong Xing, Weijie Wang, Zhengyu Zhao, Jindong Gu, Philip Torr, and Nicu Sebe. Finetune like you pretrain: Boosting zero-shot adversarial robustness in vision-language models.arXiv preprint arXiv:2604.11576, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  48. [48]

    Structure causal models and llms integration in medical visual question answering.IEEE Transactions on Medical Imaging, 44(8):3476–3489, 2025

    Zibo Xu, Qiang Li, Weizhi Nie, Weijie Wang, and Anan Liu. Structure causal models and llms integration in medical visual question answering.IEEE Transactions on Medical Imaging, 44(8):3476–3489, 2025

    2025

  49. [49]

    Transfer learningfordetectingunknownnetworkattacks

    Juan Zhao, Sachin Shetty, Jan Wei Pan, Charles Kamhoua, and Kevin Kwiat. Transfer learningfordetectingunknownnetworkattacks. EURASIP Journal on Information Security, 2019(1):1, 2019

    2019