acceptodds
Under review as a conference paper at ICLR 2027

KronDPBL: Self-Supervised Pattern–Noise Disentanglement Towards Transferable WiFi CSI Representation Learning

Abstract

WiFi Channel State Information (CSI) has become a promising sensing modality for device-free human activity recognition. However, raw CSI measurements are contaminated by hardware impairments and environmental noise, leading deep neural networks to learn spurious, noise-dependent features that generalize poorly across domains. We propose KronDPBL, a self-supervised representation learning framework that explicitly disentangles recurring signal patterns from segment-specific noise to learn robust and transferable CSI representations. KronDPBL introduces a distribution-guided cross-attention mechanism that separates persistent signal motifs from transient perturbations, together with an asymmetric attention design that retains noise information in the key branch for contrastive similarity estimation while suppressing it in the value branch to synthesize noise-free pattern representations. To exploit the intrinsic structure of CSI measurements, KronDPBL further incorporates Kronecker-sum spatial encoding and a low-rank reconstruction objective, a parameter-efficient two-axis spatial prior into the learned representations. The model is trained with dual contrastive objectives that jointly enforce pattern invariance and noise disentanglement, applied to each corpus under a strict fold-wise protocol, with encoders frozen after pretraining, and producing representations that transfer effectively across sensing tasks. Extensive experiments on six public CSI benchmark datasets demonstrate that a frozen KronDPBL encoder without end-to-end fine-tuning achieves state-of-the-art performance on 11 of 13 downstream tasks, improving accuracy by up to 14.6% over prior methods. Furthermore, KronDPBL consistently exhibits superior robustness under severe signal perturbations, outperforming existing approaches by 5.6%–23.0%.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
WiFi CSI, Self-supervised learning, Representation learning, Pattern-noise disentanglement, Contrastive learning, Kronecker factorization, Wireless sensing, Human activity recognition

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_89iXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes KronDPBL, a self-supervised framework for learning transferable representations from WiFi Channel State Information (CSI). Its central idea is to disentangle recurring signal patterns from noise using learnable distribution-guided cross-attention and an asymmetric attention mechanism that suppresses contributions from noise representations. The framework additionally incorporates Kronecker-based spatial encoding, low-rank reconstruction, and dual contrastive learning objectives. After self-supervised pretraining, the encoder is frozen and paired with lightweight classifiers for downstream tasks. Experiments across six public CSI datasets and 13 evaluation settings report state-of-the-art performance on 11 settings, supported by ablation studies, robustness tests, and cross-hardware transfer experiments.

Critical strengths
  1. Interesting methodological design for noise-robust representation learning. The asymmetric attention mechanism, which uses both pattern and noise representations as keys while suppressing noise contributions in the values, is an interesting architectural choice. Combined with Kronecker-based structural regularization, it provides a meaningful inductive bias for learning CSI representations. This is more than a straightforward application of an existing classifier to wireless measurements.

  2. Broad and convincing empirical evaluation. The paper evaluates six public datasets, four downstream task categories, and 13 experimental settings, comparing against numerous established methods. The reported improvements are substantial in several settings, and the ablations demonstrate the contributions of the main components. The robustness experiments under additive noise, subcarrier dropout, and antenna dropout further strengthen the evidence for practical effectiveness.

Critical weaknesses
  1. Insufficient evidence for the claimed pattern–noise disentanglement. Although the architecture explicitly separates pattern and noise branches, this does not necessarily mean that the learned representations correspond to meaningful signal and noise components. In particular, the noise contrastive objective treats independently corrupted views of the same CSI segment as positives and views of different segments as negatives. This encourages segment discrimination, which may preserve segment-specific signal information rather than isolate noise characteristics. The pattern–noise separation penalty does not fully resolve this ambiguity. The ablation studies support the usefulness of the architecture but do not directly establish disentanglement, which is a central scientific claim.

  2. Transferability is demonstrated only within a limited scope. Despite the emphasis on transferable representations, the framework trains separate encoders for different datasets rather than learning a unified encoder across heterogeneous CSI sources. The cross-hardware experiments are also asymmetric: CSI-Bench to WiFall achieves 78.88% accuracy, whereas the reverse direction reaches only 50.68%, close to chance. These results demonstrate some transfer capability but do not fully establish broadly transferable representations across sensing environments and hardware platforms.

Questions for authors
  1. Noise contrastive objective: In Equation (2), the noise contrastive loss treats independently corrupted views of the same CSI segment as positive pairs and views of different segments as negatives. However, this encourages the noise representations to preserve segment-specific information rather than necessarily capture noise characteristics. In particular, two segments affected by similar noise may be pushed apart simply because they originate from different recordings. What prevents the noise branch from encoding segment-specific signal patterns instead of noise? Can the authors provide empirical evidence that the learned noise representations genuinely capture noise characteristics and are disentangled from task-relevant patterns?

  2. Cross-dataset transfer: The experiments pretrain a separate encoder for each dataset, and the cross-hardware transfer results show substantial asymmetry. To what extent can the reported gains be attributed to learning generally transferable CSI representations rather than strong dataset-specific self-supervised pretraining? Can the authors clarify the intended scope of the transferability claim and explain the large performance difference between the two transfer directions?

  3. Comparison with recent self-supervised approaches: The paper discusses recent methods such as CSI-JEPA and WiFo but does not include them in the numerical evaluation. How does KronDPBL compare with these methods under matched pretraining data and downstream protocols? Such comparisons would help establish the significance of the proposed representation-learning approach relative to recent alternatives.

Reviewer_wo4fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KronDPBL, a self-supervised framework for WiFi CSI representation learning. Its core concept involves separating recurring signal patterns from segment-specific noise using a CSI Pattern Encoder (CPE) and a CSI Noise Canceller (CNC). By integrating Kronecker spatial encoding, low-rank Kronecker reconstruction, and a dual contrastive learning strategy—targeting both patterns and noise—the framework achieves more robust and transferable CSI representations.

Critical strengths
  1. The paper addresses the robustness and transferability of WiFi CSI representations, tackling a problem of significant practical relevance.
  2. The experimental evaluation is comprehensive, covering six datasets, four types of tasks, and a variety of hardware platforms and data collection conditions.
  3. The study employs a "fold-wise" pretraining protocol, ensuring that held-out recordings are strictly excluded from the pretraining phase—a rigorous experimental design.
Critical weaknesses
  1. The paper defines HH as the "pattern representation" and ZZ as the "noise representation," attempting to separate them via cosine-based separation and distinct attention/value pathways. However, this design does not inherently prove that HH and ZZ correspond to the physical phenomena of "patterns" and "noise."
  2. The framework uses two independently corrupted versions of the same segment as a positive pair, requiring their respective noise representations (ZZ) to be similar; yet, the artificial corruptions applied to these two views are inherently different.
  3. Cross-dataset transfer is successful in only one direction, which is insufficient to support a strong claim regarding "transferable representations."
Questions for authors
  1. How can it be directly demonstrated that HH truly represents the "pattern" and ZZ represents the "noise"?
  2. What is the accuracy when using HH to predict the corruption type, device ID, or environment ID?
  3. Does the robustness hold when the model encounters entirely unseen types of corruption?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.