acceptodds
Under review as a conference paper at ICLR 2027

GLCR: Global-Local Complementary Representations for Wi-Fi Probe Request Association under MAC Randomization

Abstract

Passive Wi-Fi sensing provides a promising way to support crowd monitoring by counting nearby devices when cameras are limited by occlusion or poor visibility. However, MAC address randomization breaks the connection between Wi-Fi probe requests and devices, which makes device counting unreliable. Existing methods use fingerprints derived from Information Elements (IEs) or rely on wireless measurements and frame metadata that can vary with deployment conditions. To jointly learn complementary representations of IE order and byte-level content without relying on such auxiliary information, we propose Global-Local Complementary Representation (GLCR), a self-supervised learning framework for Wi-Fi probe request association under MAC address randomization. GLCR learns representations from retained IE content and structure. Specifically, we construct a global IE sequence view that preserves IE order, and a local IE subsequence view that focuses on byte-level IE content. For the global view, we devise a hierarchical protocol encoder to capture dependencies at the byte and IE levels. For the local view, we perform salience fingerprint mining on the local subsequence to aggregate contextualized features using learned weights. We further adopt a context-gated refinement to control the local contribution to the global representation. Finally, we cluster the resulting embeddings without a predefined device count. Experiments on three public datasets and our WiDAR dataset show that GLCR outperforms all evaluated baselines, with relative gains of 3.25%–7.86% in clustering accuracy. The code is available at https://anonymous.4open.science/r/GLCR.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Wi-Fi probe request association, MAC address randomization, self-supervised learning, global-local representation learning, unsupervised clustering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_53zfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GLCR, a self-supervised framework for associating Wi-Fi probe requests under MAC address randomization. The method learns global and local representations from Information Elements (IEs) without using RSSI or timing information. The learned representations are combined and clustered using DBSCAN. The authors evaluate the method on three public datasets and a newly collected WiDAR dataset.

Critical strengths
  1. The paper proposes an approach to associating Wi-Fi probe requests without relying on RSSI or timing information. Learning representations from IE content and ordering is useful because wireless measurements can vary across environments. The self-supervised approach also avoids the need for device identity labels during training.
  2. The experimental results are promising. Table 2 shows that GLCR outperforms the evaluated baselines across all four datasets. Table 3 provides ablation studies for the proposed components. Figures 6 and 7 examine performance under different device operating modes and observation windows.
Critical weaknesses
  1. The MAC randomization behavior during data collection is not sufficiently described. Section 5.1 explains that WiDAR was collected using a passive Wi-Fi sniffer, but it is unclear whether the devices had previously connected to nearby Wi-Fi networks or were performing scans for unknown networks. This distinction matters because probe request behavior and MAC randomization can differ depending on the connection history and scanning conditions. Section 4.1 filters frames with locally administered MAC addresses, but this alone does not establish that the addresses changed during the experiment. The paper also does not clearly explain how ground-truth device identities were established for WiDAR. Since associating requests across randomized MAC addresses is the main objective, these details are important for evaluating the results.
  2. The paper does not provide a breakdown of performance across device types. Section 5.1 states that WiDAR contains 50 devices from eight vendors, including smartphones and personal computers. However, the authors do not report how many devices belong to each category, their operating system versions, or their Wi-Fi chipsets. Table 2 reports aggregate clustering accuracy, but it is unclear whether the method performs differently on iPhones, Android phones, and computers. In particular, devices with similar hardware and software configurations may generate similar IE patterns. It would be useful to understand which devices are more likely to be incorrectly associated and whether newer MAC randomization implementations affect the results.
Questions for authors
  1. Section 5.1 describes the WiDAR collection setup. Were the devices previously connected to nearby Wi-Fi networks, or were they scanning for unknown networks? Did the authors verify that each device generated multiple randomized MAC addresses? How were the ground-truth device identities established?
  2. Could the authors provide a breakdown of the WiDAR dataset by device type, model, operating system version, and Wi-Fi chipset? It would also be useful to report association performance separately for iPhones, Android phones, and computers, particularly for devices with similar hardware and software configurations.
  3. In Table 3, removing the local branch reduces ACC on WiDAR from 81.30% to 81.18%, while the decrease is much larger on WPR24 and NEST. Could the authors explain this difference? Also, could they report the computational cost of the individual components to better understand the tradeoff between model complexity and accuracy?
  4. The authors deliberately exclude RSSI and timing information to improve robustness across environments. However, these measurements could provide complementary information when multiple devices have similar IE patterns. Have the authors considered combining IE representations with RSSI or timing information? It would be useful to understand whether these additional measurements improve association accuracy, particularly for devices with similar hardware and software configurations.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 53zf (Part 1/2)

Thank you for your constructive review and suggestions. In the following, we answer the questions from you one by one.

1. Response to Ethics and Privacy

We note your concerns about the potential use of GLCR for unauthorized device tracking and the need for an ethics review. We would like to clarify that GLCR only associates probe requests originating from the same Wi-Fi device when the device uses multiple randomized MAC addresses. Neither the device's factory-assigned MAC address nor the identity of its owner is recovered or identified. Therefore, there are no privacy problems at all for the proposed technique.

In addition, existing studies, including Nozari et al. (2025b), Uras et al. (2020), Uras et al. (2022), Baccichet et al. (2024), Pintor and Atzori (2022a), and He et al. (2023), did not raise similar privacy concerns about reliable Wi-Fi device counting. Instead, they highlighted its applications, such as crowd monitoring and emergency response.

We agree that any real-world deployment should undergo a separate ethics and privacy assessment covering the scope of association, the outputs made available, data access, and retention. Sections 4.1 and 5.1 describe the preprocessing and collection procedures, respectively, where our WiDAR dataset was collected in a controlled environment with the informed consent of the device owners. The Ethical Statement describes the scope of the evaluation.

2. Response to Weakness 1 and Q1: Collection Conditions and Ground Truth

During WiDAR collection, the devices were not previously connected to nearby Wi-Fi networks, and they were scanning for unknown networks. We verified that each device generated multiple randomized MAC addresses. For example, 21 of the 25 devices (84%) generated more than 10 distinct randomized MAC addresses in the WiDAR test set. Probe requests from each ground-truth device were captured separately in a shielded environment, and all captured frames were labeled with the corresponding device identity independently of their source MAC addresses. The recordings were subsequently combined. We have clarified these procedures and the source-MAC statistics in Section 5.1.

3. Response to Weakness 2 and Q2: Device Composition and Performance

Following your suggestion, we have provided a breakdown of the available device information for the WiDAR test set in Table 9 of Appendix F. We have also added Table 10 to report association performance separately for each device group. Among the phone groups, Android phones achieve the highest ACC (85.14%) and ARI (83.64%), while HarmonyOS phones achieve the highest NMI (92.56%). Overall, GLCR demonstrates effective association performance across diverse device groups, including groups containing devices with similar model and OS configurations.

Table 10. Association performance by device group on WiDAR.

Device groupACC (%)NMI (%)ARI (%)
iPhone77.1789.2176.45
Android phones85.1491.0083.64
HarmonyOS phones78.0092.5679.33
other devices85.9394.2195.39

4. Response to Q3: Local Branch Contribution and Computational Cost

The different ACC reductions across datasets are mainly attributed to variations in the discriminative information provided by the selected IE subsequences. The local branch encodes these subsequences to complement the complete IE-sequence representation learned by the global branch. On WiDAR, the selected IEs contain fewer device-discriminative cues, resulting in a smaller contribution from the local branch. In contrast, on WPR24 and NEST, the selected IEs provide more useful complementary information, leading to larger performance gains. We have clarified this dataset-dependent behavior in the Ablation Study.

We have added Table 7 in Appendix F to report the computational costs of GLCR and its component ablations, including parameter counts, FLOPs, and inference latency. The local Transformer and SFM comprise 6.71K parameters and require 0.65M FLOPs per frame. Removing the local branch and its associated CGR reduces inference latency from 1.93 to 1.22 ms per frame. Together with the dataset-specific accuracy gains in Table 3, these results quantify the tradeoff between association performance and computational cost.

Table 7. Network-level costs of GLCR-Base and its ablations.

ConfigurationParams (K)FLOPs (M)Latency (ms)
GLCR-Base28.668.341.93
w/o Local branch21.127.681.22
w/o Global branch10.820.650.57
w/o IEHA25.603.861.41
w/o SFM28.518.331.89
w/o CGR27.838.341.84
w/o Gate27.868.341.85
AuthorsOpenReview ↗

Response to Reviewer 53zf (Part 2/2)

5. Response to Q4: RSSI and Timing Information

Following your suggestion, we evaluated whether combining RSSI and timing improves the association accuracy. The results show that neither feature improved the association performance of GLCR in our experiments. On WiDAR, ACC decreased from 81.30% (IE-only GLCR) to 76.51% (+RSSI), 80.78% (+Time), and 78.59% (+RSSI+Time), with corresponding decreases in NMI and ARI (Appendix F, Table 11). Furthermore, we compared existing approaches that combined RSSI and timing in Table 2 of Section 5, such as Cappuccino, CrowdWatch, and TC-Hybrid. The results show that GLCR consistently outperforms these methods in ACC, NMI, and ARI across all four datasets, despite relying solely on IE content and structure.

These results are attributed to the instability of RSSI and timing under different device behaviors and measurement conditions. Our measurements showed that RSSI values for the same source MAC address fluctuate between −90 and −34 dBm, even at a fixed device-to-sniffer distance. Timing is likewise affected by scanning behavior, device state, missed frames, and irregular transmissions. Although more sophisticated fusion strategies may benefit devices with similar IE characteristics, the evaluated configurations did not improve overall association performance. We have clarified the motivation for IE-only association in the Introduction.

Table 11. Clustering results (%).

ConfigurationACC ↑NMI ↑ARI ↑
GLCR81.3092.2683.36
GLCR + RSSI76.5188.9480.02
GLCR + Time80.7891.5182.24
GLCR + RSSI + Time78.5988.4680.20
Reviewer_53zfOpenReview ↗

Response to Authors and Final Assessment

Thank you for the detailed responses and additional experiments. The new gate ablation addresses my main concern, while the additional baseline comparisons and device-count results strengthen the paper.

I appreciate the revisions and will maintain my current rating.

AuthorsOpenReview ↗

Ethics and Privacy Concerns

We appreciate your valuable suggestions. In the future, we will continue to improve our work by taking your professional suggestions and questions into consideration.

On the other hand, we noticed that you marked the `Flag For Ethics Review’ and had some concerns about the potential use of the proposed technique GLCR for unauthorized WiFi device tracking. We would like to clarify that the technique only associates probe requests from the same Wi-Fi device when the device uses multiple randomized MAC addresses. Neither the device's factory-assigned MAC address nor the identity of its owner is recovered or identified. Therefore, there are no privacy problems at all for the proposed technique. In addition, existing studies, including Nozari et al. (2025b), Uras et al. (2020), Uras et al. (2022), Baccichet et al. (2024), Pintor and Atzori (2022a), and He et al. (2023), did not raise similar privacy concerns about reliable Wi-Fi device counting. Instead, they highlighted its applications, such as crowd monitoring and emergency response. Our WiDAR dataset was collected in a controlled environment with the informed consent of the device owners. The revised Ethical Statement describes the collection conditions, data minimization, limits on association scope, and controlled artifact release.

We hope our responses have adequately addressed your privacy concerns about the proposed technique. If so, we would sincerely appreciate it if you may remove the Flag For Ethics Review.

Please feel free to let us know if you have any further questions or concerns. Thank you again for your time and constructive feedback.

Reviewer_EkXpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduce a self-supervised representation method for Wi-Fi probe request association. The key problem to address is how to cluster the same-device frames together, given that modern mobile devices usually adopt MAC address randomization for privacy. The core idea is to use IE order and byte-level content as complementary inputs to generate a "Global-Local Complementary Representation (GLCR)".

Critical strengths
  • It is an interesting direction to sense the crowd with WiFi signals instead of camera images.
  • The problem targeted in this paper - WiFi probe request association - is a core challenge in WiFi sensing.
Critical weaknesses
  • The core contribution seems to be about WiFi signal scheme understanding, instead of a novel representation learning algorithm. The self-supervised algorithm seems to be a combination of existing techniques.
  • The novelty over existing WiFi probe request clustering approaches is not clearly explained. Many design choices are difficult to understand for readers without wireless communication background.
Questions for authors
  • What is the key challenge in clustering or learning representations for WiFi probe requests? Is the challenge generic for other types of data?
  • I'm a bit confused about the comparison with related work. The paper says "we study joint representation learning from ordered IE sequences and selected byte-level content without these auxiliary cues". Does that mean that you included new features (e.g. byte-level content) for clustering? If so, why is this feature useful, and why didn't prior work use it? I also find it difficult to understand the separation of "global view" and "local view". What is your definition of global and local, and why could such a separation help learning in principle?
4 comments
AuthorsOpenReview ↗

Response to Reviewer EkXp (Part 1/2)

Thank you for your constructive review and suggestions. In the following, we answer the questions from you one by one.

1. Response to Ethics and Privacy

We note your concern that the proposed technique, GLCR, seems to be opposed to Wi-Fi MAC address randomization, which is intended to protect user privacy. We would like to clarify that GLCR only associates probe requests originating from the same Wi-Fi device when the device uses multiple randomized MAC addresses. It neither recovers the device's factory-assigned MAC address nor identifies its owner. Therefore, there are no privacy problems at all for the proposed technique.

In addition, existing studies, including Nozari et al. (2025b), Uras et al. (2020, 2022), Baccichet et al. (2024), Pintor and Atzori (2022a), and He et al. (2023), did not raise similar privacy concerns regarding reliable Wi-Fi device counting. Instead, these studies highlighted its applications, such as crowd monitoring and emergency response.

We agree that any real-world deployment should undergo a separate ethics and privacy assessment covering the scope of association, the outputs made available, data access, and retention. Sections 4.1 and 5.1 describe the preprocessing and collection procedures, respectively, where our WiDAR dataset was collected in a controlled environment with the informed consent of the device owners. The Ethical Statement describes the scope of the evaluation.

2. Response to Weakness 1: Methodological Contribution

The core methodological contribution of GLCR lies in explicitly incorporating Wi-Fi protocol structure into representation construction, encoding, and optimization, rather than simply combining existing self-supervised learning components. Specifically, Element ID membership guides hierarchical encoding and IE compactness regularization, while unlabeled Value-field length statistics guide the construction of complementary local views. We have added comparisons with three existing representation-learning methods—RoMAE, ET-BERT, and VICReg. GLCR achieves the highest ACC, NMI, and ARI across all four datasets, outperforming the strongest of these three representation-learning baselines in ACC by 6.88%, 7.09%, 8.34%, and 2.66%, respectively (Section 5.2, Table 2).

Table 2 (excerpt). Clustering performance of GLCR and existing representation learning methods on IE sequences. Each entry reports ACC / NMI / ARI (%).

MethodWPR22WPR24NESTWiDAR
RoMAE (Zivanovic et al., 2025)75.68 / 82.21 / 78.6674.69 / 90.14 / 77.3965.63 / 82.22 / 55.2168.09 / 87.04 / 67.13
ET-BERT (Lin et al., 2022)76.63 / 83.55 / 79.5677.91 / 91.98 / 79.4179.94 / 90.23 / 79.0379.19 / 91.11 / 80.76
VICReg (Bardes et al., 2022)76.69 / 82.00 / 77.8964.32 / 85.48 / 54.9363.29 / 78.00 / 46.4859.89 / 82.40 / 48.39
GLCR (Ours)81.97 / 86.13 / 80.3583.43 / 93.49 / 87.9786.61 / 92.96 / 84.0781.30 / 92.26 / 83.36

The specific methodological contributions are as follows:

a. Protocol-Aware Hierarchical Encoding (IEHA). Unlike generic byte-level Transformers, IEHA groups contextualized bytes by Element ID, models interactions among IE groups, and injects IE-level context back into byte representations. The same grouping defines IE compactness regularization, incorporating protocol structure into both encoding and the learning objective (Section 4.2).

b. Protocol-Guided Global–Local Views (IAP). Rather than selecting arbitrary tokens, IAP constructs the local view by selecting complete IEs using unlabeled Value-field length statistics while preserving their original relative order. This provides a structured local view complementary to the complete global IE sequence (Section 4.1).

c. Context-Dependent Local Aggregation (SFM). SFM learns context-dependent token weights within the IAP-selected IE subsequence. Its contribution lies in integrating protocol-guided selection with contextual aggregation, rather than introducing a new attention mechanism (Section 4.3).

d. Context-Gated Refinement (CGR). CGR uses a residual form of feature-wise gated fusion that preserves an ungated path for the global embedding. A gate conditioned on both embeddings scales the local contribution before addition to the global embedding and LayerNorm (Section 4.4).

Ablation Evidence. Table 3 further supports these designs. Removing IEHA reduces WiDAR ACC by 6.53%, removing the local branch and its associated refinement reduces WPR24 ACC by 20.44%, and removing the learned gate while retaining LayerNorm reduces WPR24 ACC by 14.72%.

We have revised the Introduction and Section 2 to clarify these methodological contributions and their distinctions from existing representation-learning approaches.

AuthorsOpenReview ↗

Response to Reviewer EkXp (Part 2/2)

3. Response to Weakness 2 and Q2: Related Work and View Definitions

Thank you for pointing out the need to clarify our distinction from prior work and the motivation for the global–local design. Regarding input features, GLCR does not introduce new raw input features; prior methods have also used IE content, including Value bytes. Our contribution lies in incorporating existing IE content and protocol structure into representation learning. Regarding the view definitions, the global view contains the complete preprocessed IE sequence, while the local view contains an adaptively selected subsequence of complete IEs. Both preserve byte content and relative IE order. Regarding the motivation, this separation allows the local branch to focus on selected IE content, potentially reducing interference from less informative fields, while the global branch retains complete IE-sequence context.

IE Features and Distinction from Prior Work. Each IE contains an Element ID, Length, and Value. Differences in Value bytes, even within the same IE type, can provide useful device association cues. For example, Cappuccino combines IE information with auxiliary cues such as RSSI and sequence numbers for pairwise frame association, followed by minimum-cost flow. In contrast, GLCR learns per-frame representations from IE content and structure, then clusters them without a predefined device count.

We have clarified these points in the Introduction, Section 2, and Section 4.

4. Response to Q1: Key Challenge and Generality

Key Challenge. The key challenge is that different devices can generate similar IE patterns, while requests from the same device may contain different IE content. Thus, GLCR must learn device-discriminative representations despite this ambiguity, without device-identity labels or a predefined device count.

Generality. These challenges reflect general difficulties in representation learning and clustering, including high within-class variation, between-class similarity, and an unknown number of clusters. However, Wi-Fi probe requests introduce additional challenges: MAC randomization removes reliable address-based identity cues, while the remaining information consists of typed, variable-length IE fields whose content and order must be modeled.

We have clarified these challenges in the Introduction, with the task formulation provided in Section 3.

I'm sorry but I still cannot grasp the algorithmic innovation of this work even after careful reading. Without proper understanding of the contribution, I cannot change my rating, but I'd ask the AC to take my low expertise in WiFi communication into final consideration.

AuthorsOpenReview ↗

Thank you for your candid feedback and consideration

Dear Reviewer EkXp,

Thank you very much for taking the time to read our paper and response carefully, and for your candid and fair feedback.

To ensure clarity for both you and the Area Chair, we briefly summarize our algorithmic novelty and how it differs from existing approaches from a general representation learning perspective:

Difference from prior Wi-Fi association methods: Instead of relying on heuristic byte templates or environment-sensitive metadata (such as RSSI and timing), GLCR formulates device association as self-supervised representation learning solely on hierarchical Type-Length-Value (TLV) byte sequences.

Difference from standard sequence representation models (e.g., RoMAE, ET-BERT, VICReg): Rather than treating a packet as a flat token sequence, GLCR explicitly exploits element-membership structure through three tailored designs:

1. Hierarchical Protocol Encoder (HPE/IEHA): performs token-to-element hierarchical aggregation, cross-element interaction, and Element-ID compactness regularization rather than flat self-attention alone;

2. Adaptive Subsequence View & Salience Mining (IAP/SFM): constructs a complementary local view from Element IDs with high Value-field length variability and learns frame-specific token salience weights;

3. Context-Gated Refinement (CGR): uses residual feature-wise gating to control the local contribution while preserving an ungated global path.

Both our ablation study (Table 3) and direct comparisons against these sequence models (Table 2) confirm that each structure-aware component contributes substantial gains without relying on domain-specific wireless metadata.

On the other hand, we noticed that you marked the `Flag For Ethics Review’ and had some concerns about the potential use of the proposed technique GLCR for unauthorized WiFi device tracking. We would like to clarify that the technique only associates probe requests from the same Wi-Fi device when the device uses multiple randomized MAC addresses. Neither the device's factory-assigned MAC address nor the identity of its owner is recovered or identified. Therefore, there are no privacy problems at all for the proposed technique. In addition, existing studies, including Nozari et al. (2025b), Uras et al. (2020), Uras et al. (2022), Baccichet et al. (2024), Pintor and Atzori (2022a), and He et al. (2023), did not raise similar privacy concerns about reliable Wi-Fi device counting. Instead, they highlighted its applications, such as crowd monitoring and emergency response. Our WiDAR dataset was collected in a controlled environment with the informed consent of the device owners. The revised Ethical Statement describes the collection conditions, data minimization, limits on association scope, and controlled artifact release.

We hope our responses have adequately addressed your concerns regarding both the algorithmic novelty and privacy safeguards of our work. If so, we would sincerely appreciate it if you could remove the ethics-review flag and reconsider your overall assessment accordingly.

Please feel free to let us know if you have any further questions or concerns. Thank you again for your time and constructive feedback.

Best regards,

The Authors

Reviewer_JvP4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents GLCR, a self-supervised framework for associating Wi-Fi probe requests under MAC address randomization. The main idea is to learn complementary representations from the ordered Information Element (IE) sequence and selected byte-level IE content, without relying on auxiliary information such as RSSI or timestamps. The framework uses a hierarchical encoder for the global IE sequence, an adaptive selection strategy and salience-weighted aggregation for the local view, and a gated fusion module to combine the two representations. The resulting embeddings are clustered using DBSCAN without requiring the number of devices in advance. The authors evaluate the method on three public datasets (WPR22, WPR24, and NEST) and their newly collected WiDAR dataset. GLCR is trained on unlabeled WiDAR data and evaluated using a frozen encoder on all four datasets. The results show consistent improvements over six baseline methods, supported by ablation studies and additional analyses of device operating modes, observation windows, and model configurations.

Critical strengths
  1. Strong and consistent experimental results. GLCR achieves the best ACC, NMI, and ARI across all four datasets among the evaluated methods. The baselines cover several types of approaches, including IE-based clustering, methods using multiple association cues, and self-supervised learning. The improvements are consistent across datasets and metrics, which provides convincing evidence of the method's effectiveness.
  2. A meaningful cross-dataset evaluation. I particularly appreciate that the authors train the encoder only on the unlabeled WiDAR training set and evaluate it on three public datasets without retraining. The WiDAR training and test splits also contain disjoint sets of devices. This is a stronger evaluation than training and testing a separate model on each dataset, and it provides useful evidence that the learned representations transfer to unseen devices and independently collected data. The use of IE content and structure alone, without relying on RSSI, timestamps, or other auxiliary measurements, is also a useful feature of the method.
  3. Well-motivated design and solid ablation studies. The global-local design makes sense for this task. The ordered IE sequence and the fine-grained content of individual IEs provide different kinds of information, and the proposed architecture attempts to capture both. The ablation studies are reasonably comprehensive. Removing either the global or local branch leads to performance degradation, particularly on WPR24 and NEST, supporting the complementary roles of the two representations. The experiments also examine the contributions of IEHA, SFM, and CGR. In addition, the analyses of observation windows, device modes, and model configurations make the evaluation more complete.
Critical weaknesses
  1. The methodological novelty could be better clarified. While the architecture is tailored to Wi-Fi probe requests, several of its general ideas, such as complementary representation learning, attention-based aggregation, and gated feature fusion, have been explored in previous work. For example, DREAM (ICLR 2026) uses dual-branch representations with dynamic gated fusion for clustering, although its data setting and technical design are different. I see the main contribution of GLCR as its use of protocol-aware IE representations rather than the general idea of combining two branches with a gate. The hierarchical modeling of byte-to-IE relationships and the construction of the local IE view are interesting, but I would have liked a clearer explanation of what makes these designs distinct from existing representation learning approaches. This would help better establish the paper's methodological contribution.
  2. The CGR ablation could be more carefully controlled. Table 3 shows a large performance drop when CGR is removed. However, Appendix F indicates that the w/o CGR variant removes both the gating mechanism and LayerNorm, replacing the entire module with simple averaging. As a result, it is difficult to tell how much of the improvement comes specifically from the feature-wise gate rather than the normalization operation. Since the learned embeddings are used for distance-based clustering, this distinction seems relevant. An additional comparison that keeps LayerNorm while removing the gate would make the ablation more informative. This does not undermine the effectiveness of the complete CGR module, but it would help explain which part of the design is responsible for the improvement.
Questions for authors

Q1. In the CGR ablation, could the authors include a comparison using LayerNorm(z_g + z_l), which removes the feature-wise gate but retains normalization? This would help isolate the contribution of the gating mechanism. Q2. Could the authors clarify which aspects of the proposed global-local representation design are the main methodological contributions compared with existing complementary representation learning and gated-fusion approaches? Q3. Since GLCR performs clustering without requiring the number of devices in advance, could the authors report how closely the estimated number of clusters matches the actual number of devices on each dataset? A comparison between the ground-truth device count and the predicted cluster count would help assess the method's effectiveness for device counting, in addition to the reported clustering metrics.

1 comment
AuthorsOpenReview ↗

Response to Reviewer JvP4

Thank you for your constructive review and suggestions. We answer your questions below.

1. Response to Weakness 1 and Q2: Methodological Novelty

The main methodological contribution of GLCR is a protocol-aware representation-learning architecture, rather than the general use of complementary branches, attention, or gated fusion. Specifically, Element ID membership guides both hierarchical encoding and IE compactness regularization, while Value-field length statistics from unlabeled frames guide the construction of a complementary local view that is selectively integrated into the global representation. Unlike DREAM (ICLR 2026), which combines semantic and inter-sample structural representations, GLCR explicitly models byte-to-IE and cross-IE relationships within each probe request. We have revised the Introduction, Section 2, and Section 4 to better clarify the main methodological contributions and their distinctions from existing representation learning approaches.

The specific methodological distinctions are as follows:

a. Protocol-Aware Hierarchical Encoding (IEHA). Unlike generic byte-level Transformers, IEHA groups contextualized bytes by Element ID, models interactions among IE groups, and injects IE-level context back into byte representations. The same grouping also defines IE compactness regularization, incorporating protocol structure into both encoding and the learning objective (Section 4.2).

b. Protocol-Guided Global–Local Views (IAP). The global view retains the complete IE sequence, while IAP selects complete IEs using unlabeled Value-field length statistics and preserves their original relative order. This constructs a protocol-guided local view rather than an arbitrary token subset (Section 4.1).

c. Context-Dependent Local Aggregation (SFM). SFM assigns context-dependent weights to encoded bytes within the IAP-selected subsequence. Its contribution lies in integrating protocol-guided selection with contextual aggregation, rather than introducing a new attention mechanism (Section 4.3).

d. Context-Gated Refinement (CGR). CGR uses a residual form of feature-wise gated fusion that preserves an ungated path for the global embedding. A gate conditioned on both embeddings scales the local contribution before addition to the global embedding and LayerNorm (Section 4.4).

Experimental Evidence. We have added comparisons with RoMAE, ET-BERT, and VICReg in Section 5.2 (Table 2). GLCR achieves the highest ACC, NMI, and ARI across all four datasets, outperforming the strongest of these three representation-learning baselines in ACC by 6.88%, 7.09%, 8.34%, and 2.66%, respectively. Ablation results (Table 3) further support the effectiveness of individual components: removing IEHA reduces WiDAR ACC by 6.53%, replacing SFM with mean pooling reduces WPR24 ACC by 4.54%, and removing the learned gate while retaining LayerNorm reduces WPR24 ACC by 14.72%.

2. Response to Weakness 2 and Q1: Gate Ablation with LayerNorm

We have added the requested w/o Gate ablation, which retains LayerNorm while removing the learned feature-wise gate. The results show that removing the gate reduces ACC by 3.09%, 14.72%, 17.57%, and 18.25% on four datasets, respectively (Table 3). This controlled comparison demonstrates that the learned gate improves association performance beyond the effect of LayerNorm alone. Specifically, w/o Gate uses z_c=LayerNorm(z_g+z_l), whereas w/o CGR removes both gating and normalization, replacing the module with simple averaging. We have clarified these configurations in the revised Ablation Study and Appendix F.

Selected ACC results from Table 3: Effect of feature-wise gating with LayerNorm retained.

MethodWPR22WPR24NESTWiDAR
w/o Gate79.4471.1571.3966.46
GLCR81.9783.4386.6181.30
Relative ACC drop (%)3.0914.7217.5718.25

3. Response to Q3: Estimated Device Counts

We have added Table 4 (Section 5.2) to directly compare the estimated and ground-truth device counts among frames retained after filtering in Section 4.1. On four datasets, GLCR estimates an average of 17.60, 12.80, 12.60, and 21.80 devices, compared with the ground-truth counts of 17, 16, 13, and 25, respectively. The corresponding mean absolute errors (MAEs) are 1.40, 3.60, 2.00, and 3.20 devices. The absolute differences between the mean estimated counts and the ground-truth counts do not exceed 3.20 devices across the four datasets. In particular, the mean estimates differ from the ground truth by only 0.60 and 0.40 devices on WPR22 and NEST, respectively.

Table 4. Ground-truth and GLCR-estimated device counts.

DatasetKKK^‾\overline{\hat K}ΔK\Delta KMAE
WPR221717.60+0.601.40
WPR241612.80−3.203.60
NEST1312.60−0.402.00
WiDAR2521.80−3.203.20

Comments on the paper

2 comments
AuthorsOpenReview ↗

Ethics and privacy clarification for GLCR

Dear TPC Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

Two of the three reviewers had privacy concerns about the proposed technique GLCR in this paper, which seems to be opposed to Wi-Fi MAC address randomization. We would like to clarify that the proposed technique GLCR only identifies probe requests from the same Wi-Fi device when the device uses multiple randomized MAC addresses. The real identity information about the device, such as its factory-assigned MAC address and the device owner, is neither recovered nor identified. Therefore, there are no privacy problems at all for the proposed technique.

In addition, existing studies, Nozari et al. (2025b), Uras et al. (2020), Uras et al. (2022), Baccichet et al. (2024), Pintor & Atzori (2022a), and He et al. (2023), did not show similar privacy concerns about the reliable Wi-Fi device counting problem studied in this paper. In contrast, they believed that reliable Wi-Fi device counting has many applications, such as crowd monitoring and emergency response.

We agree that any real-world deployment should undergo a separate ethics and privacy assessment covering the scope of association, the outputs made available, data access, and retention.

In Sections 4.1 and 5.1, we described the preprocessing and collection procedures, respectively, where our WiDAR dataset was collected in a controlled environment with the informed consent of the device owners. The Ethical Statement describes the scope of the evaluation.

Thank you for your consideration.

The Authors

AuthorsOpenReview ↗

Summary of Revisions, Contributions, and Reviewer Discussions

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank the Area Chairs and all reviewers for their time, effort, and constructive feedback throughout the review process, which have helped us significantly strengthen the paper. We have carefully addressed all comments and summarize our core contributions, experimental evidence, and reviewer discussions below to assist with the final assessment.

1. Core Methodological Contribution & Key Strengths

Rather than relying on brittle heuristic byte templates or environment-sensitive wireless metadata (such as RSSI or timestamps), GLCR formulates device association as structure-aware self-supervised representation learning solely on hierarchical Type-Length-Value (TLV) byte sequences. Unlike flat token-stream models, GLCR incorporates protocol element-membership structure through three coordinated components:

  • Hierarchical Protocol Encoding & Regularization (Section 4.2): Aggregates contextualized byte features by Element ID (IE), models cross-element interactions, injects group context back into byte representations, and enforces IE compactness regularization—going beyond flat token-level self-attention alone.

  • Adaptive Local-View Construction & Salience Mining (Sections 4.1 & 4.3): Replaces heuristic random masking with an ordered subsequence of complete IEs selected via unlabeled Value-field length variability statistics, coupled with learned frame-specific token salience weights.

  • Context-Gated Refinement (Section 4.4): Uses residual feature-wise gating to adaptively regulate local-branch contributions while preserving an ungated global path.

2. Key Revisions & Empirical Evidence

All revisions are highlighted in the updated manuscript and provide strong empirical support across four datasets:

  • State-of-the-Art Performance Across All Benchmarks (Table 2, Section 5.2): Added comparisons against strong general and traffic sequence representation baselines (RoMAE, ET-BERT, and VICReg) alongside six task-specific baselines. GLCR achieves the highest ACC, NMI, and ARI across all four datasets, delivering relative ACC improvements of 2.66%–6.88% over the strongest baseline without physical-layer wireless metadata.

  • Clear Component-Wise Necessity (Table 3, Section 5.2): Added comprehensive ablations across aggregation, pooling, fusion, and branches. With LayerNorm retained, GLCR achieves 83.43% ACC on WPR24 versus 71.15% for fixed summation (+12.28% gain from adaptive context gating).

  • Accurate Downstream Device Counting (Table 4, Section 5.2): Added estimated counts, signed errors, and MAEs (with ground truth defined over retained post-preprocessing frames), achieving mean errors within 3.20 devices overall and within 0.60 devices on WPR22 and NEST.

  • Additional Practical Analyses (Appendix F): Added device composition, device-group performance, computational efficiency, and RSSI/timing fusion experiments.

3. Reviewer Discussions and Ethical Safeguards

  • Reviewer 53zf confirmed that the new gate ablation addressed their main concern and that the added baseline and device-count evaluations strengthened the paper, maintaining their positive rating.

  • Reviewer JvP4 gave a positive initial evaluation, and we provided detailed responses and manuscript revisions addressing all of their comments.

  • Reviewer EkXp maintained their rating, noting remaining uncertainty about the algorithmic innovation and candidly asking the Area Chair to consider their limited background in Wi-Fi communication in the final decision. We appreciate their transparent feedback; accordingly, we further clarified our algorithmic contributions in the revised manuscript (Sections 1 and 4) and provided a domain-agnostic summary from a representation-learning perspective in our final reply, inviting the Area Chairs to weigh these concrete structural designs alongside the empirical evidence in Tables 2 and 3.

  • Ethical Safeguards: The expanded Ethical Statement documents informed consent, controlled data collection, data minimization, strict within-collection association boundaries, and controlled artifact release, directly addressing potential privacy risks from linking randomized MAC addresses.

Thank you again for your time, oversight, and careful consideration of our work.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.