Multi-Surrogate-Teacher Assistance for Representation Alignment in Fingerprint-based Indoor Localization

Son Minh Nguyen, Tran Duy Linh, Duc Viet Le, Paul J. M. Havinga

Abstract

Despite remarkable progress in knowledge transfer across visual and textual domains, extending these achievements to indoor localization, particularly for learning transferable representations among Received Signal Strength (RSS) fingerprint datasets, remains a challenge. This is due to inherent discrepancies among these RSS datasets, largely including variations in building structure, the input number and disposition of WiFi anchors. Accordingly, specialized networks, which were deprived of the ability to discern transferable representations, readily incorporate environment-sensitive clues into the learning process, hence limiting their potential when applied to specific RSS datasets. In this work, we propose a plug-and-play (PnP) framework of knowledge transfer, facilitating the exploitation of transferable representations for specialized networks directly on target RSS datasets through two main phases. Initially, we design an Expert Training phase, which features multiple surrogate generative teachers, all serving as a global adapter that homogenizes the input disparities among independent source RSS datasets while preserving their unique characteristics. In a subsequent Expert Distilling phase, we continue introducing a triplet of underlying constraints that requires minimizing the differences in essential knowledge between the specialized network and surrogate teachers through refining its representation learning on the target dataset. This process implicitly fosters a representational alignment in such a way that is less sensitive to specific environmental dynamics. Extensive experiments conducted on three benchmark WiFi RSS fingerprint datasets underscore the effectiveness of the framework that significantly exerts the full potential of specialized networks in localization.

Introduction

Indoor localization often relies on Received Signal Strength (RSS) fingerprinting—a method of mapping WiFi or Bluetooth signal strengths to specific coordinates. While deep learning has advanced this field, models are often brittle and environment-specific. A model trained for one building typically fails in another due to "non-conformity": variations in building layout, the number of WiFi anchors, and environmental noise. Traditional transfer learning struggles because these feature spaces are too disparate, leaving state-of-the-art models isolated in their original training environments.

Methodology

The proposed framework addresses these challenges through a two-phase "Plug-and-Play" (PnP) scheme designed to foster environment-insensitive representations.

The proposed PnP framework consisting of Expert Training (a) and Expert Distilling (b) phases.

Phase 1: Expert Training

Rather than transferring raw data, the framework trains surrogate generative teachers using Wasserstein GANs (WGANs) to model the representation spaces of various source datasets. These teachers act as global adapters, homogenizing disparate inputs while preserving the unique characteristics of the source environments.

Phase 2: Expert Distilling

The specialized "student" model is trained on the target dataset under the guidance of three unsupervised constraints:

  • Angular Similarity ($J_{Sim}$): Aligns the direction of learned feature vectors with those of the teachers.
  • Cross-Mutual Information ($J_{MI}$): Maximizes shared essential knowledge while filtering out environment-specific noise.
  • Functional Information ($J_{FI}$): Uses the Lipschitz constant (estimated via the Power Iteration Method) to match the functional sensitivity and robustness of the networks.

Experimental Results

The framework was evaluated on three benchmark datasets: UJIIndoorLoc (large-scale campus), UTS (controlled area), and Tampere (high anchor sparsity). The "fully enhanced" models (denoted with +++) consistently outperformed their standalone counterparts across various architectures, including DNNs, CNNs, and Transformers.

Localization performance comparison (MAE and Percentiles) across three benchmark datasets.

Key findings include:

  • Accuracy Improvement: Mean Absolute Error (MAE) was reduced by 11% to 20% across different models. For instance, in UJIIndoorLoc, a standard DNN's error dropped by 2.06 meters.
  • Reliability: The probability of localization error being under 10 meters increased from approximately 50-60% to 62-75%.
  • Stability: Significant improvements were observed in the 75th and 95th percentile error rates, indicating a reduction in extreme miscalculations.

Conclusion

This work demonstrates that distilling knowledge through surrogate teachers effectively sidesteps privacy and data heterogeneity issues in indoor localization. By prioritizing transferable features, the framework filters out environmental "clutter" that typically leads to overfitting. While highly effective for most architectures, future work is needed to adapt the Functional Information ($J_{FI}$) constraint for the non-smooth attention mechanisms in Transformers, which currently require its omission to maintain performance stability.

Summary updated

Abstract

Despite remarkable progress in knowledge transfer across visual and textual domains, extending these achievements to indoor localization, particularly for learning transferable representations among Received Signal Strength (RSS) fingerprint datasets, remains a challenge. This is due to inherent discrepancies among these RSS datasets, largely including variations in building structure, the input number and disposition of WiFi anchors. Accordingly, specialized networks, which were deprived of the ability to discern transferable representations, readily incorporate environment-sensitive clues into the learning process, hence limiting their potential when applied to specific RSS datasets. In this work, we propose a plug-and-play (PnP) framework of knowledge transfer, facilitating the exploitation of transferable representations for specialized networks directly on target RSS datasets through two main phases. Initially, we design an Expert Training phase, which features multiple surrogate generative teachers, all serving as a global adapter that homogenizes the input disparities among independent source RSS datasets while preserving their unique characteristics. In a subsequent Expert Distilling phase, we continue introducing a triplet of underlying constraints that requires minimizing the differences in essential knowledge between the specialized network and surrogate teachers through refining its representation learning on the target dataset. This process implicitly fosters a representational alignment in such a way that is less sensitive to specific environmental dynamics. Extensive experiments conducted on three benchmark WiFi RSS fingerprint datasets underscore the effectiveness of the framework that significantly exerts the full potential of specialized networks in localization.
BibTeX
@inproceedings{nguyen2025multi,
  author = {Son Minh Nguyen and Tran Duy Linh and Duc Viet Le and Paul J. M. Havinga},
  title = {{Multi-Surrogate-Teacher Assistance for Representation Alignment in Fingerprint-based Indoor Localization}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
  url = {https://openreview.net/forum?id=Sx1FRETfeT},
}
Figure