Physical layer key generation based on SOM and K-Means clustering quantization for TDD-OFDM systems over unstable channels

Weijie Tan, Qiangqiang Gao, Zeyu Li, Zhiquan Liu, Chunguo Li, Gang XuJournal of King Saud University - Computer and Information Sciences 2025 · DOI

TL;DR

This paper introduces a robust physical layer key generation scheme for wireless systems that struggle with "unstable" channels. By combining a deep learning-based Channel Reciprocity Learning Network (CRLN) to align channel features with a hybrid SOM-KMeans clustering algorithm for quantization, the authors significantly reduce key disagreement rates and ensure high-entropy, random keys even in noisy environments.

What Problem Does This Paper Address?

In secure wireless communication, "Alice" (the sender) and "Bob" (the receiver) can generate secret cryptographic keys by measuring the unique, random characteristics of the wireless channel between them. This relies on channel reciprocity: the idea that the signal path from Alice to Bob is identical to the path from Bob to Alice.

In the real world, this reciprocity is a myth. Hardware asymmetries in radio circuits, environmental noise, and multipath interference create discrepancies between Alice’s and Bob’s measurements. These differences lead to a high Key Disagreement Rate (KDR), meaning Alice and Bob end up with different keys, rendering the system useless without heavy (and often slow) error correction. Existing methods often fail in "unstable" channels where signal-to-noise ratios (SNR) are low or channel conditions fluctuate rapidly.

Methods

The authors propose a two-stage pipeline to bridge the reciprocity gap and generate reliable keys:

  1. Channel Reciprocity Learning Network (CRLN): The authors use a feedforward neural network to perform "feature mapping." Since Alice cannot see Bob's exact channel state, the CRLN is trained to map Alice’s local observations to a version that closely resembles Bob’s. This effectively "denoises" the hardware and environmental discrepancies before key generation even begins.
  2. SOM-KMeans Quantization: Once the features are aligned, they must be converted into binary bits (0s and 1s). Standard methods use fixed thresholds, which are brittle. Instead, the authors use a hybrid unsupervised learning approach:
  3. Self-Organizing Maps (SOM): These map high-dimensional channel data onto a 2D grid, preserving the topological structure and reducing dimensionality.
  4. K-Means Clustering: This algorithm then groups the SOM-processed data into clusters.
  5. Gray Coding: Finally, these clusters are mapped to bit sequences. Because Gray coding ensures that adjacent clusters differ by only one bit, it minimizes the impact of small estimation errors.

Datasets

The study relies on a dataset generated via the MATLAB WLAN Toolbox, simulating a TDD-OFDM system under a Rician fading model (which accounts for both line-of-sight and scattered signal components).

  • Training Set: 10,000 samples of Channel State Information (CSI).
  • Testing Set: 2,000 samples.
  • Environment: The simulation mimics various propagation models, ranging from "Mild" (indoor with minimal obstacles) to "Extremely Strong" (underground parking lots or tunnels).

Key Insights and Ideas

  • Deep Learning as a "Correction Fluid": Rather than trying to build perfect hardware, the authors use the CRLN to mathematically "align" two different views of the same channel. This treats reciprocity violation as a regression problem that a neural network is well-suited to solve.
  • Unsupervised Adaptability: By using SOM and K-Means, the quantization process becomes adaptive. Instead of Alice and Bob agreeing on a hard numerical threshold (which might shift with noise), they agree on the structure of the data clusters.
  • Security via Decorrelation: The paper confirms that while Alice and Bob’s keys become more similar, a passive eavesdropper ("Eve") located more than a few wavelengths away sees a completely different, uncorrelated channel, keeping the keys secure.

Key Results

  • Drastic KDR Reduction: In comparative tests, the proposed SOM-KMeans method achieved a Key Disagreement Rate of 4.3%, a massive improvement over traditional direct quantization, which sat at 21.7%.
  • Eavesdropper Resistance: The success rate for an eavesdropper (Eve) to guess the key dropped to a negligible 0.05%.
  • Randomness Confirmed: The generated keys passed the rigorous NIST Statistical Test Suite, proving they are sufficiently random for cryptographic use.
  • Performance at High SNR: The system achieves a superior Key Generation Rate (KGR) as the signal quality improves, outperforming Auto-Encoder and LSTM-based quantization methods.

Limitations and Future Work

  • Computational Latency: The primary trade-off is speed. The SOM-KMeans approach has a latency of 9.8 ms, compared to just 2.7 ms for direct quantization. While acceptable for many IoT applications, it might be a bottleneck for ultra-low-latency tactical systems.
  • Offline Training: The CRLN currently requires an offline training phase with pre-collected data. In a truly dynamic, "zero-touch" deployment, the system would need a way to perform online or federated learning to adapt to entirely new environments.
  • Real-World Hardware: The results are currently based on high-fidelity simulations. The authors suggest that moving to real-world hardware testbeds (like USRP platforms) is the necessary next step to validate these gains against actual radio frequency (RF) impairments.

Summary updated

Abstract

In TDD-OFDM systems, key generation relies on channel reciprocity, which assumes that the channel characteristics of the transmit and receive channels are identical. However, practical conditions such as hardware asymmetry, environmental noise, and multipath effects often disrupt this reciprocity, negatively impacting the accuracy and consistency of key generation. To address these issues, this paper proposes a novel key generation scheme based on Self-Organizing Maps (SOM) and K-Means quantization algorithm for TDD-OFDM system under unstable channel. The proposed approach uses Channel Reciprocity Learning Network (CRLN) for feature mapping to enhance the similarity of channel characteristics between the base station (Alice) and the legitimate user (Bob), effectively reducing the key disagreement rate. The physical layer key is derived from two sets of similar channel features, and the initial key is generated using the SOM-KMeans clustering algorithm combined with Gray coding. Experimental results demonstrate that this proposed method significantly lowers the key disagreement rate in unstable channels and achieves a superior key generation rate at high Signal-to-Noise Ratios (SNR) compared to other quantization techniques. Furthermore, the initial keys produced successfully pass the NIST randomness tests.
BibTeX
@article{tan2025physical,
  author = {Weijie Tan and Qiangqiang Gao and Zeyu Li and Zhiquan Liu and Chunguo Li and Gang Xu},
  title = {{Physical layer key generation based on SOM and K-Means clustering quantization for TDD-OFDM systems over unstable channels}},
  journal = {Journal of King Saud University - Computer and Information Sciences},
  year = {2025},
  doi = {10.1007/s44443-025-00245-6},
}
Figure