PaperlayerView exact revision on arXiv

Source-linked AI summary

Channel-Informed Neural Network for Physical Layer Key Generation

Jose Angel Sanchez Viloria, George Sklivanitis, Dimitris Pados, Elizabeth Serena Bentley

arXiv:2609.16341v1cs.LGeess.SP

TL;DR

Edge devices may need lightweight key establishment with limited access to centralized infrastructure, while existing PKG work gives comparatively less attention to raw-IQ features grounded in multipath structure. This paper trains a channel-informed network to generate reciprocal binary features, and reports lower disagreement for Alice–Bob observations than Eve-related observations alongside higher key diversity with ray-traced augmentation.

  • Problem

    Edge devices need lightweight shared-key establishment despite limited access to centralized infrastructure, and comparatively less PKG work explicitly grounds raw-IQ features in multipath-channel structure.

    Source: §I · p. 1

  • Method

    A multi-task channel-informed network learns reciprocity-preserving binary features from raw IQ while predicting an auxiliary channel representation, using structured probing and ray-traced training data.

    Source: §I · pp. 1–2, B. Neural Network Architecture · p. 3

  • Results

    Across evaluated scenarios, reciprocal Alice–Bob observations had lower bit disagreement than Eve-related observations; ray-traced augmentation raised unique-key rates to 0.94, 0.99, and 0.99.

    Source: B. Experimental Results · p. 5, §VI · p. 6

  • Takeaways & Limitations

    The results support channel-informed representation learning as a potential approach to decentralized wireless key establishment, with key diversity and reconciliation reliability remaining a design tradeoff.

    Source: §I · p. 2, §VI · p. 5

Abstract

from arXiv · show

Physical-layer key generation (PKG) enables wireless devices to establish shared keys from reciprocal channel observations without directly exchanging the key. This capability is attractive for edge networks, where distributed and resource-constrained devices may require lightweight key establishment with limited access to centralized infrastructure. We introduce a channel-informed neural network for PKG that derives binary key features directly from received IQ measurements while explicitly grounding the learned representation in the underlying multipath channel. The proposed multi-task recurrent neural network jointly learns reciprocity-preserving binary features and an auxiliary channel estimate using a training objective that combines deep metric learning with channel-informed supervision. Structured channel sounding enables channel estimation from over-the-air measurements, while Sionna-RT ray tracing is used to augment training with additional propagation conditions. We evaluate the framework using indoor and outdoor software-defined-radio measurements collected on the POWDER radio testbed. Across all evaluated scenarios, the proposed model produces lower bit disagreement for reciprocal Alice-Bob observations than for Eve-related observations. Ray-traced data augmentation substantially improves key diversity, increasing the unique-key rate to 0.94, 0.99, and 0.99 across the indoor and two outdoor scenarios, respectively. Successfully reconciled channel-informed keys pass the selected NIST randomness tests prior to SHA-3 privacy amplification. The results demonstrate the potential of channel-informed representation learning for decentralized wireless key establishment while highlighting an important tradeoff between key diversity and reconciliation reliability.

I. INTRODUCTION

The introduction motivates lightweight, decentralized key establishment for edge devices and identifies a gap in PKG methods that learn directly from raw IQ while preserving multipath-channel structure. It presents a channel-informed learning framework and reports improved key diversity alongside a tradeoff with reconciliation success.

I. INTRODUCTION

  • Edge devices may need lightweight key establishment despite limited or intermittent access to centralized infrastructure.Secure key establishment is foundational for edge communications, but centralized key-management entities can add computational, communication, and operational overhead.Source: §I · p. 1
  • Prior PKG work often uses precomputed channel descriptors or optimizes individual pipeline stages, leaving less attention to raw-IQ features explicitly constrained by multipath structure.The introduction notes data-driven approaches for practical key agreement, including methods for learned features and hardware impairments –.Source: §I · p. 1
  • The proposed framework combines structured channel probing, auxiliary channel-estimation supervision, and ray-traced propagation data to ground learned binary features in the wireless environment.It builds on prior AI-assisted PKG work and supports a pipeline comprising feature extraction, reconciliation, privacy amplification, and dynamic data encryption.Source: §I · pp. 1–2
  • Across evaluated indoor and outdoor settings, channel-informed training separates reciprocal legitimate-channel observations from eavesdropper observations while exposing a tradeoff between key diversity and reconciliation success.The introduction also reports that ray-traced augmentation increases unique-key rates relative to the baseline.Source: §I · p. 2

II. SYSTEM MODEL

The system model uses reciprocal Alice–Bob channel probes and a passive Eve, with over-the-air estimates and ray tracing representing propagation channels. It assumes legitimate links are approximately reciprocal while spatial separation makes Eve’s channel less correlated.

II. SYSTEM MODEL

  • Alice and Bob exchange known probes in TDD within the channel coherence time, while passive Eve monitors without modifying them.Their forward and reverse observations are expected to be highly correlated, but noise, hardware impairments, and synchronization offsets can make them differ,.Source: §II · p. 2
  • The probe is a BPSK-modulated GLFSR code sequence, and the received signal model includes multipath attenuation, path delays, and complex Gaussian noise.The path count is P, with h_p and τ_p denoting path attenuation and delay, respectively.Source: §II · p. 2
  • Because the probing sequence is known at the receiver, the system obtains a least-squares channel estimate from the received probe samples and a Toeplitz convolution matrix.The channel vector contains P taps; the ray tracer also computes channel impulse responses for simulated propagation paths.Source: §II · p. 2
  • Ray-traced and estimated Alice–Bob and Bob–Alice channel impulse responses are expected to be approximately reciprocal, while Eve’s channel is expected to be less correlated when spatially separated.These channel responses ground neural feature extractors operating directly on received IQ spectrograms.Source: §II · p. 2

III. CHANNEL-INFORMED LEARNING PKG FRAMEWORK

The framework aims to generate binary keys directly from IQ spectrograms while grounding reciprocity-preserving features in the physical channel. Its workflow covers data collection, model training, and testing.

III. CHANNEL-INFORMED LEARNING PKG FRAMEWORK

  • The framework targets reciprocity-preserving binary keys derived directly from IQ spectrograms and grounded in the wireless channel between trusted nodes.Figure 1 presents the end-to-end workflow for data collection, training, and testing.Source: §III · p. 2
  • The workflow connects data collection to model training and subsequent testing of the proposed PKG framework.The stated goal is key generation for a pair of trusted network nodes.Source: §III · p. 2
  • Received channel-probe IQ spectrograms are the stated input to the framework’s feature-generation objective.The framework seeks binary outputs that preserve channel reciprocity.Source: §III · p. 2

A. Signal Pre-processing

Preprocessing normalizes each received probe before converting its complex IQ samples into a real-valued time–frequency representation. Overlapping Hanning-windowed segments and an STFT capture that structure for the neural feature extractor.

A. Signal Pre-processing

  • RMS normalization is applied to each received IQ probe to reduce received-power variation across links and probing rounds.A small positive constant is included for numerical stability.Source: A. Signal Pre-processing · p. 3
  • The normalized sequence is divided into overlapping segments, and a Hanning window is applied to reduce spectral leakage before the discrete Fourier transform.The segments use window length NF and hop size R.Source: A. Signal Pre-processing · p. 3
  • The STFT forms a time–frequency matrix whose real and imaginary components are separated to provide real-valued neural-network inputs.The representation is formed from the complex STFT coefficients.Source: A. Signal Pre-processing · p. 3

B. Neural Network Architecture

The channel-informed feature extractor uses a recurrent encoder with shared representations for binary key features and auxiliary channel estimation. Its architecture is designed to retain temporal context from received-probe spectrograms.

B. Neural Network Architecture

  • A bidirectional GRU processes the spectrogram sequence, aggregating context across received-probe time-frequency slices.The input has M = 16 slices with NF = 32 features per step; the GRU has 384 hidden units per direction.Source: B. Neural Network Architecture · p. 3
  • The encoder refines and pools GRU outputs into a fixed-length, L2-normalized latent vector shared by both task-specific heads.The dense normalization blocks, global average pooling, and projection produce z ∈ R^W.Source: B. Neural Network Architecture · p. 3, Figure 2 · p. 3
  • The binary-feature head applies mean-centered sign quantization to produce the features used for key generation.The selected key length is W = 128, and a straight-through estimator supports end-to-end training through the sign operation.Source: B. Neural Network Architecture · p. 3
  • A second head estimates complex channel taps from the continuous embedding during training, while key generation requires only the binary-feature head.The channel head outputs P = 16 complex taps in the described implementation.Source: B. Neural Network Architecture · p. 3

C. Channel-Informed Deep Metric Learning

Training combines a metric-learning objective for separating reciprocal and eavesdropper observations with channel-informed supervision. The channel constraint guides representations to preserve multipath-channel information rather than relying only on relative embedding distances.

C. Channel-Informed Deep Metric Learning

  • The overall training objective combines data-driven metric learning with a channel-informed loss.The channel weight λch controls the channel constraint's contribution.Source: C. Channel-Informed Deep Metric Learning · p. 4
  • The metric-learning objective encourages reciprocal Alice–Bob and Bob–Alice embeddings to be closer than Alice–Eve and Bob–Eve embeddings.A third term further separates the legitimate reciprocal pair from the eavesdropper pair.Source: C. Channel-Informed Deep Metric Learning · p. 4
  • The channel-informed loss penalizes disagreement between estimated channel taps and reference taps for each received probe.Reference taps come from the GLFSR-based probe's LS estimate or the Sionna-RT ray tracer.Source: C. Channel-Informed Deep Metric Learning · p. 4
  • Unlike relative-distance constraints alone, channel-informed supervision guides the representation to preserve information about the multipath channel.The channel loss uses normalized mean-squared error over Alice–Bob, Bob–Alice, Alice–Eve, and Bob–Eve links.Source: C. Channel-Informed Deep Metric Learning · pp. 3–4

IV. RECONCILIATION AND PRIVACY AMPLIFICATION

After binary features are generated, Alice and Bob reconcile residual disagreements with Reed–Solomon coding and apply privacy amplification to the reconciled sequence. The method also describes how parity information is exposed to Eve and how code rate is selected.

IV. RECONCILIATION AND PRIVACY AMPLIFICATION

  • Alice sends Reed–Solomon parity symbols, which Bob uses with his own sequence to correct symbol errors and recover Alice's sequence.The decoder can correct up to ⌊(N − K)/2⌋ unknown symbol errors.Source: §IV · p. 4
  • Because Eve can observe public parity symbols, the code rate balances Alice–Bob correction capability against side information available to Eve.Eve is assumed able to attempt reconciliation using the same code.Source: §IV · p. 4
  • After successful reconciliation, Alice and Bob hash the reconciled bit sequence for privacy amplification.The POWDER outdoor SDR and Sionna-RT digital-twin setups are used for training augmentation, testing, and evaluation of the protocol.Source: §IV · p. 4, Figure 3 · p. 4

A. Experimental Setup

The evaluation combines indoor and outdoor POWDER software-defined-radio measurements with ray-traced data used to augment training. The setup includes controlled synthetic channel-probing simulations and reserved outdoor OTA data for testing.

A. Experimental Setup

  • Ray tracing generated 7,500 synthetic spectrogram and channel quadruplets across 500 virtual Alice–Bob–Eve triplets and 15 SNR values from 5 dB to 30 dB.The ray-traced channel taps drove GNU Radio software-in-the-loop channel-probing simulations on a 10×10 spatial grid.Source: A. Experimental Setup · p. 4
  • Indoor OTA data comprise 1,000 quadruplets from 10 topologies, while outdoor OTA measurements cover two topologies using three POWDER radios.The indoor setup used four B210 and four X310 SDRs separated by approximately 6.4 m; the outdoor radios were at USTAR, EBC, and Guest House.Source: A. Experimental Setup · p. 4
  • The dataset uses 256-symbol BPSK GLFSR probes, 16 channel taps, a 3.5 GHz carrier, and 1 MHz receiver sampling.The spectrogram inputs are 32 × 16 × 2, and the RNN produces 128-bit features and a 16-tap complex channel estimate.Source: A. Experimental Setup · pp. 4–5
  • Indoor and simulation data were split into training, validation, and testing sets, while the full 200-quadruplet outdoor OTA dataset was reserved for testing.The split was 70%, 15%, and 15%, respectively; Fig. 4 covers BDR distributions across indoor and outdoor Alice–Bob–Eve topology configurations and three models.Source: A. Experimental Setup · p. 5, Figure 4 · p. 5

B. Experimental Results

Across the OTA scenarios, reciprocal Alice–Bob observations show lower disagreement than Eve-related observations, while ray-traced augmentation substantially increases key diversity. Reconciliation results expose a diversity–agreement tradeoff, and successfully reconciled keys pass the selected NIST tests before privacy amplification.

B. Experimental Results

  • The BDR comparison uses 150 indoor, 95 outdoor 1, and 92 outdoor 2 probe exchanges, with Alice–Bob, Alice–Eve, and Bob–Eve pairs evaluated.After successful reconciliation, Alice and Bob hash the reconciled binary feature vector for privacy amplification.Source: B. Experimental Results · p. 5
  • Across all three OTA scenarios, Alice–Bob BDR consistently indicates more similar binary vectors for reciprocal legitimate channels than for eavesdropper channels.The comparison covers the RNN, CI-RNN trained on indoor OTA data, and CI-RNN trained on indoor OTA plus Sionna-RT data.Source: B. Experimental Results · p. 5
  • Unique-key rates with indoor-OTA and Sionna-RT training reach 0.94, 0.99, and 0.99 for indoor, outdoor 1, and outdoor 2, respectively.CI-RNN trained only on indoor OTA data yields 0.37, 0.63, and 0.57; the RNN yields 0.21, 0.29, and 0.33.Source: B. Experimental Results · p. 5
  • Lower-rate RS codes raise Alice–Bob reconciliation rates but also allow Eve to reconcile, while proposed CI-RNN reconciliation rates are lower than with its lower key uniqueness.Reconciliation uses RS(24, 16) and RS(32, 16) over 8-bit symbols; success means Bob's decoded sequence matches Alice's binary vector.Source: B. Experimental Results · p. 5
  • 100% of successfully reconciled CI-RNN keys meet the evaluated NIST randomness tests, whereas generated keys in fail the Random Excursion Variant test.The nine tests are applied before privacy amplification.Source: B. Experimental Results · p. 5

VI. CONCLUSIONS AND FUTURE WORK

The paper presents a channel-informed neural framework for generating PKG features from raw IQ measurements, evaluated on POWDER OTA measurements with Sionna-RT training augmentation. Results highlight improved key diversity alongside an agreement–diversity tradeoff and identify broader validation as future work.

VI. CONCLUSIONS AND FUTURE WORK

  • The multi-task framework jointly learns reciprocity-preserving binary features and an auxiliary channel representation, grounding key features in measurable propagation structure.It combines structured channel probing, multipath-channel supervision, and ray-traced data augmentation in a unified learning pipeline.Source: §VI · p. 5
  • Across evaluated scenarios, the channel-informed RNN produces lower bit disagreement for reciprocal Alice–Bob observations than for Eve-related observations, and successfully reconciled keys pass selected NIST tests.Evaluation uses indoor and outdoor POWDER SDR measurements together with Sionna-RT-generated propagation scenarios for training augmentation.Source: §VI · pp. 5–6
  • Greater key diversity coincides with increased residual Alice–Bob disagreement and reduced reconciliation success under the evaluated Reed–Solomon configurations.Future work includes reconciliation-aware training, adaptive quantization and error correction, alternative channel-informed losses, and validation under mobility, interference, larger topologies, and active attacks.Source: §VI · p. 6
Loading 2609.16341v1…