PaperlayerView exact revision on arXiv

Source-linked AI summary

gr-PHYSEC: Real-time Channel-based Key Generation for Physical Layer Secure Wireless Communications

Jose Angel Sanchez Viloria, George Sklivanitis, Dimitris Pados

arXiv:2609.16375v1cs.CR

TL;DR

Wireless key generation and exchange remain challenging for resource-constrained devices in infrastructure-free ad-hoc networks. gr-PHYSEC embeds neural channel-feature extraction and key reconciliation in GNU Radio; experiments report 2–15% typical bit disagreement and 80% key-generation success with RS(255,128).

  • Problem

    Symmetric-key generation and exchange remain difficult for resource-constrained devices communicating over wireless links in infrastructure-free ad-hoc networks.

    Source: §1 · p. 1

  • Method

    gr-PHYSEC embeds a trained CNN-based channel-feature extractor in GNU Radio and uses quantization, Reed-Solomon reconciliation, and hashing to generate keys from bidirectional probes.

    Source: §1 · pp. 1–2, §2 · p. 2

  • Results

    Across 100 real-time generations, Alice–Bob bit disagreement was typically 2–15%, and RS(255,128) reconciliation achieved an 80% key-generation success rate.

    Source: §5.1 · p. 4

  • Takeaways & Limitations

    gr-PHYSEC provides a practical, reproducible real-time key-generation pipeline implemented in GNU Radio and tested on low-cost SDRs with edge-AI platforms.

    Source: §7 · p. 5

    The authors identify increasing key-generation success, reducing generation time, and assessing security against Eve's position and sophistication as future work.

    Source: §7 · p. 5

Abstract

from arXiv · show

Securing wireless communication against eavesdropping is critical, particularly in dynamic and decentralized environments. We present gr-PHYSEC, a new GNU Radio out-of-tree (OOT) module for real-time physical-layer key generation. Unlike traditional key generation that relies on pre-shared secrets or computational complexity, our approach derives symmetric keys from the wireless channel's inherent randomness. We embed a trained neural network within GNU Radio to extract channel features between trusted parties (Alice and Bob) during probe exchanges. These features are quantized into binary keys, reconciled via Reed-Solomon encoding, and further secured with SHA-512 hashing. The generated keys are then directly used to encrypt data. Real-world experiments at the FAU CAAI connected robotics testbed using ADALM Pluto software-defined radios and NVIDIA Jetson Orin validate the approach with ground robotic platforms. Results demonstrate low key disagreement rates and strong randomness, as verified by the NIST test suite for random and pseudorandom number generators for cryptographic applications. This integration showcases how GNU Radio can support real-time AI-driven security solutions, pushing the boundaries of software-defined secure communication. The source code for this project is available at: https://github.com/C2A2-at-Florida-Atlantic-University/gr-PHYSEC

1. Introduction

gr-PHYSEC addresses real-time symmetric-key generation for decentralized wireless links by embedding a channel-based key-generation pipeline in GNU Radio. Its OOT blocks combine neural feature extraction, key reconciliation, and privacy amplification for deployment on SDRs and Jetson platforms.

1. Introduction

  • Prior physical-layer key-generation systems use channel-derived RSS, CSI, or CIR features and reconcile disagreements, while GNU Radio implementations have also supported LoRa key generation and other PHY-security techniques.The introduction frames symmetric-key generation and exchange as a challenge for resource-constrained devices in infrastructure-free ad-hoc networks.Source: §1 · p. 1
  • gr-PHYSEC demonstrates real-time physical-layer key generation for on-the-fly encryption in peer-to-peer ad-hoc wireless communications.The authors implement and test GNU Radio OOT signal-processing blocks for key generation between Alice and Bob.Source: §1 · p. 1
  • A trained CNN maps probe spectrograms to channel embeddings, which are quantized into bit vectors, reconciled with Reed-Solomon coding, and hashed into 512-bit keys.The hashing step is described as reducing information leaked to Eve during reconciliation.Source: §1 · p. 2
  • The open-source OOT project includes blocks for probing, STFT spectrogram generation, ONNX inference, quantization, Reed-Solomon reconciliation, and SHA-512 privacy amplification.It also provides control state machines, ready-to-run flowgraphs, a pretrained model, and Docker containers for Jetson and ADALM-Pluto deployment.Source: §1 · p. 2

2. Secret Key Generation Protocol

The protocol turns reciprocal wireless probes into a shared key through feature extraction, quantization, error reconciliation, and privacy amplification. Its CNN is trained to preserve Alice–Bob similarity while separating their embeddings from Eve's.

2. Secret Key Generation Protocol

  • Alice and Bob probe bidirectionally, extract and quantize channel embeddings, reconcile disagreements with transmitted Reed-Solomon parity, and apply privacy amplification to the resulting key.The protocol's overview identifies privacy-amplification precedents in the cited literature.Source: §2 · p. 2

2.1. Channel-based feature extraction

  • A 15-layer ResNet maps 31×256 log-power spectrograms to 512-element embeddings, trained with quadruplet loss to minimize Alice–Bob distance and maximize each party's distance from Eve.Each node buffers 8192 IQ samples; the spectrogram uses a 256-sample Hamming window with 50% overlap.Source: §2.1 · p. 2

2.2. Key generation

  • With RS(255,128), Bob can detect up to 127 erroneous symbols or correct up to 63 unknown-location errors; SHA-3-512 maps the reconciled vector to a 512-bit key.The text assumes Alice's parity symbols reach Bob without errors and states that a one-bit difference in Eve's reconciled vector yields a completely different hash output.Source: §2.2 · p. 3

3. gr-PHYSEC

gr-PHYSEC implements the key-generation pipeline as GNU Radio OOT blocks, with separate processing for signal features, reconciliation, privacy amplification, and coordination between SDRs.

3. gr-PHYSEC

  • The implementation comprises six GNU Radio OOT blocks for key generation plus additional blocks for probing and information exchange, with flowgraphs for Alice and Bob.Source: §3 · p. 3
  • The feature path preprocesses IQ samples into STFT log-power spectrograms, runs ONNX ResNet inference, and quantizes embeddings using each embedding's arithmetic mean as the threshold.These blocks generate the initial secret messages at Alice and Bob.Source: §3 · p. 3
  • Alice's RS(255,128) encoder generates parity bits, and Bob's decoder uses them with his quantized features to reconcile to Alice's message.A Python/C++ wrapper connects the GNU Radio parity-generation and reconciliation blocks to a Reed-Solomon library.Source: §3 · p. 3
  • The privacy-amplification block hashes reconciled messages and lets users select SHA-3 variants and output key sizes of 128, 256, or 512 bits.Source: §3 · p. 3
  • A lightweight state machine sequences bidirectional probe requests and transmissions, parity exchange, and reconciliation acknowledgments using PMT and ZMQ Pub/Sub messages.An Rx gate controls when received samples pass through the PlutoSDR source at either node.Source: §3 · p. 3

4. Experimental Setup

The experiments use paired SDR-equipped robotic agents and a third SDR acting as a passive eavesdropper in an indoor FAU CAAI lab. Nodes run the system on Jetson platforms while collecting protocol-stage data and key-generation metrics.

4. Experimental Setup

  • The indoor FAU CAAI setup collected bidirectional Alice–Bob probes while a third SDR listened as Eve, and deployed the CNN on Jetson Orin and Jetson Nano devices.The CNN was trained on an NVIDIA A100 GPU and exported to ONNX for deployment.Source: §4 · p. 3
  • Each robotic agent mounted a Jetson Orin interfaced with an ADALM-Pluto SDR; the PlutoSDRs used a 2.485 GHz carrier and 1 MSps sample rate while one node moved no more than ~5 m.A KPI monitor collected IQ samples, spectrograms, quantized channel feature vectors, and reconciliation success to calculate bit disagreement rate, key-generation success rate, and key-generation time.Source: §4 · p. 4

5. Results

Real-time trials reported 80% key-generation success with RS(255, 128), while motion and orientation increased bit disagreement. NIST tests assessed generated keys, and average generation time was approximately 1299 ms.

5.1. Bit disagreement and reconciliation

  • 80% key-generation success was achieved with RS(255, 128) across 100 real-time generations.The success rate is reported as resulting from reconciliation in most attempts.Source: §5.1 · p. 4, Figure 4 · p. 4
  • BDR typically ranged from 2–15%, increasing to approximately 56% with motion and orientation; the representative median was approximately 12.51%.The effects of motion and orientation after epoch 60 appear in Figures 4 and 5.Source: §5.1 · p. 4, Figure 5 · p. 4
  • RS(255, 128) reconciled most attempts, while shorter RS(191,128) codes reduced parity overhead but also reduced success rate.The reported BDR remained within RS correction capabilities.Source: §5.1 · p. 4

5.2. Key randomness (NIST STS)

  • NIST assessment averaged p-values across nine representative tests and calculated key pass rates over 100 independently generated keys.Each test treated a reconciled, privacy-amplified cryptokey as one binary sequence; scores are summarized in Table 1.Source: §5.2 · p. 4, Table 1 · p. 5

5.3. Latency and key generation rate

  • Average key-generation time was approximately 1299 ms on the Jetson Orin embedded setup.The estimate includes approximately 400 ms of wait time for bidirectional probe exchanges.Source: §5.3 · p. 5

6. Reproducibility

The project provides source code, installation guidance, example flowgraphs, and a Docker container to support deployment and reproducibility.

6. Reproducibility

  • The gr-PHYSEC OOT blocks, example flowgraphs, and ONNX model are available in the project repository.Installation notes cover the library and ONNX Runtime on Jetson platforms.Source: §6 · p. 5
  • Example flowgraphs configure Alice and Bob SDR nodes, while the KPI monitor reports probe exchanges, BDR, key-generation success rate, and generation time.The monitor is included with the ready-to-run flowgraphs.Source: §6 · p. 5, Figure 7 · p. 5
  • A Docker container includes ARM processor requirements and block-level parameter defaults; blocks can be swapped, including for different channel-probing waveforms.Source: §6 · p. 5

7. Conclusion and Future Work

gr-PHYSEC is presented as a practical, reproducible real-time key-generation pipeline implemented in GNU Radio and tested on mobile ground-robot links using low-cost COTS SDRs and edge-AI platforms. Future work targets higher success rates, lower generation times, and security assessment against Eve.

7. Conclusion and Future Work

  • The paper presents gr-PHYSEC as a practical, reproducible real-time physical-layer key-generation pipeline implemented in GNU Radio.Source: §7 · p. 5
  • Experiments used mobile ground-robot links with low-cost COTS SDRs interfaced to edge-AI inference platforms.The links experienced a variety of channel conditions due to node mobility.Source: §7 · p. 5
  • Future work will target increased key-generation success and reduced generation time through MAC-layer design, and assess security against Eve's position and sophistication.Source: §7 · p. 5
Loading 2609.16375v1…