PaperlayerView exact revision on arXiv

Source-linked AI summary

FedSwitch: Federated Region Classification From Wireless Channel Measurements

Mattia Piana, Stefano Rini, Stefano Tomasin

arXiv:2610.03523v1eess.SP

TL;DR

Wireless IoT region classification can use local channel measurements for physical-layer security, but heterogeneous local datasets can make a single global model suboptimal. FedSwitch combines collaborative federated training with loss-triggered local fine-tuning. Evaluations on 3GPP-compliant synthetic channel measurements report better error probability than state-of-the-art federated and conventional non-federated baselines.

  • Problem

    Heterogeneous local channel datasets can make a single global model suboptimal for classifying transmitters into regions.

    Source: §I · p. 1

  • Method

    FedSwitch combines collaborative federated training with local fine-tuning, switching when local training loss stagnates.

    Source: §I · p. 2, §V · p. 6

  • Results

    FedSwitch outperforms state-of-the-art federated and conventional non-federated baselines in error probability on realistic 3GPP-compliant channel models.

    Source: §VII · p. 11

  • Takeaways & Limitations

    Federated training of channel-based region classifiers can keep raw measurements local while reducing their exposure as location-dependent radio fingerprints.

    Source: §I · p. 2

Abstract

from arXiv · show

Wireless Internet-of-Things (IoT) networks can leverage locally observed channel state information (CSI) at base-stations (BSs) to perform region classification, i.e., infer the region of origin of a transmitter (TX), thereby enabling channel-based authentication. However, centralizing these measurements incurs substantial communication overhead and requires sharing location-dependent radio fingerprints. To address these difficulties, we propose an edge-native federated learning (FL) framework where BSs collaboratively train a shared classifier via a parameter server (PS) without exposing raw data, thereby supporting physical-layer security tasks like spoofing detection. To overcome the statistical heterogeneity arising from diverse BS locations and propagation conditions, we introduce FedSwitch, a personalized FL algorithm that automatically transitions edge nodes from collaborative training to local fine-tuning upon loss stagnation. We provide theoretical convergence guarantees for the proposed collaborative training phase and derive analytical bounds linking classification errors to statistical discrepancies. Extensive evaluations on synthetic channel measurements demonstrate that FedSwitch significantly improves region-classification accuracy over standard FL baselines.

I. INTRODUCTION

The paper formulates channel-based region classification as a way for base stations to infer a transmitter’s region from local channel observations for physical-layer security. It proposes FedSwitch, which combines federated learning with local fine-tuning to address heterogeneous channel data without sharing raw measurements.

II. PROBLEM SETTING: CHANNEL-BASED REGION CLASSIFICATION

  • Base stations classify transmitters into predefined regions from local channel observations, supporting location-aware authentication and rogue-transmitter detection.The setting considers geographically distributed base stations observing location-dependent uplink channels in a scarce-data regime.Source: §I · p. 1, §II · pp. 2–3

I. INTRODUCTION

  • Federated learning lets base stations collaboratively train a classifier through a parameter server while keeping raw channel measurements local.This avoids centralizing location-dependent radio fingerprints and reduces exposure of sensitive channel data.Source: §I · p. 2
  • FedSwitch combines collaborative training with local fine-tuning when a base station’s training loss stagnates, adapting to heterogeneous propagation conditions.The approach is intended to balance shared learning with local specialization.Source: §I · pp. 1–2
  • FedSwitch improves over standard federated baselines in heterogeneous deployments using 3GPP-compliant synthetic channel measurements.The paper also presents theoretical convergence analysis and bounds relating classification performance to discrepancies among local channel distributions.Source: §I · p. 2

II. PROBLEM SETTING: CHANNEL-BASED REGION CLASSIFICATION

  • The channel model assumes single-antenna transmitters and base stations equipped with M antennas; observations contain location-dependent information from the propagation environment.The task assigns each received observation to one of the predefined regions.Source: §II · pp. 2–3

A. Performance Evaluation

The section defines how CRC performance is evaluated and motivates keeping channel measurements local. The decision rule’s miss and false-alarm probabilities are summarized across base stations and true regions, with the DET curve linked to physical-layer verification.

A. Performance Evaluation

  • CRC performance is evaluated using miss-detection and false-alarm probabilities under each true region, summarized across base stations and regions.The average DET is represented as a tuple across the hypotheses and base stations.Source: A. Performance Evaluation · p. 3
  • A DET curve can support lightweight verification by checking whether the inferred region matches the transmitter’s claimed or authorized region.A mismatch provides evidence that the received signal may be spatially inconsistent with the authorized region.Source: A. Performance Evaluation · p. 3

III. LEARNING FRAMEWORK: PERSONALIZED FEDERATED LEARNING

  • Base stations train classifiers to infer a received signal’s region while raw channel measurements remain local, using federated collaboration through a parameter server.Centralization is undesirable because channel observations are location-dependent radio fingerprints.Source: §III · p. 3, A. Local Datasets and Classification… · p. 4

A. Local Datasets and Classification Loss

  • Each base station’s classifier estimates probabilities over the predefined region labels from its local channel observations.The classifier is trained using local datasets and a multi-class cross-entropy loss.Source: A. Local Datasets and Classification… · pp. 3–4

B. Federated Learning Preliminaries

FedAvg coordinates local training through a parameter server, but heterogeneous channel conditions can make a single global model suboptimal for individual base stations. FedSwitch addresses this with adaptive switching from collaboration to local specialization.

B. Federated Learning Preliminaries

  • The parameter server broadcasts a global model, aggregates participating base stations’ local updates, and produces the standard FedAvg update.Each selected base station trains on its local dataset before aggregation.Source: B. Federated Learning Preliminaries · p. 4
  • Different base-station positions and propagation conditions create heterogeneous local risks whose minimizers may differ, making one global model suboptimal for some stations.Source: B. Federated Learning Preliminaries · p. 4

C. Motivation for Personalized Federated Learning

  • Because global updates can become less representative as training progresses, the goal is to retain early collaboration while adapting each model to its station’s channel statistics and region distribution.The personalized objective also retains information shared during federated training.Source: C. Motivation for Personalized… · p. 4

IV. PROPOSED SOLUTION: FEDSWITCH

  • FedSwitch first trains collaboratively, then uses local loss stagnation to switch toward fine-tuning, with the final global model initializing personalization once all stations are ready.A station is marked ready when its windowed loss improvement falls below a tolerance; ready stations continue FedAvg until all are ready.Source: §IV · p. 5

A. Collaborative FedAvg Phase

The collaborative phase uses local momentum updates and parameter-server aggregation to train a shared model. Once personalization begins, each base station blends its local model with the global model and ultimately uses a station-specific classifier.

A. Collaborative FedAvg Phase

  • Each participating base station initializes from the broadcast global model, performs E local momentum-SGD steps, and sends its updated model for FedAvg aggregation.The updated global model is then evaluated on each station’s local loss to apply the switching rule.Source: A. Collaborative FedAvg Phase · p. 5

B. Personalized Fine-Tuning Phase

  • Personalization begins after the collaborative stage, with each base station initializing its personalized model from the final global model.The transition waits for all base stations to be ready.Source: B. Personalized Fine-Tuning Phase · p. 5
  • During fine-tuning, each station combines its locally adapted model with the global model, progressively reducing the global contribution as local adaptation grows.The resulting procedure returns one personalized classifier per base station for local region-of-origin inference.Source: B. Personalized Fine-Tuning Phase · p. 5

V. PERFORMANCE ANALYSIS

The analysis studies FedSwitch’s switching behavior and collaborative-phase optimality gap under stated loss and gradient assumptions. It links loss stagnation to an eventual switch and describes how heterogeneity and local-update choices affect the timing and cost.

  • A base station becomes ready for fine-tuning when its average loss reduction over a sliding window falls below a predefined tolerance, and the phase begins once all stations are ready.This criterion is intended to avoid switching prematurely or wasting communication rounds through excessive switching.Source: §V · p. 6
  • Theorem 5.1 is titled “Collaborative-phase optimality gap” and is stated under Assumptions 1 to 4, with L, µ, σk, and G defined therein.The adopted assumptions are smoothness, strong convexity, bounded variance, and bounded second moment.Source: §V · p. 6

1. Choose

The analysis bounds the binary-classification performance difference between FedAvg and the optimal Neyman–Pearson test in terms of distribution mismatch and test thresholds. It also relates heterogeneity to switching behavior and limits the scope of its convergence analysis.

  • The expected windowed loss improvement decays as O(1/e), so any positive tolerance triggers the switching condition in finite time almost surely.Source: §1 · p. 7
  • Greater heterogeneity delays the algorithm’s switch, while more frequent local updates increase energy cost.Source: §1 · p. 7
  • The strongly convex analysis is a tractable model, not a characterization of the nonconvex neural-network objective; finite-window switching guarantees require separate analysis.Source: §1 · p. 7
  • Theorem 5.2 bounds the missed-detection probability difference between FedAvg and the Neyman–Pearson test using distribution mismatch and false-alarm-specific thresholds.The accompanying discussion states that greater distribution mismatch worsens performance and that the bound also depends on the thresholds.Source: §1 · p. 7

VI. NUMERICAL RESULTS

The simulations compare FedSwitch with centralized, local-only, and federated baselines under heterogeneous multi-BS conditions, using a scenario with few training samples per BS.

VI. NUMERICAL RESULTS

  • The Global baseline pools data as an i.i.d. ideal, Single trains locally, and FedAvg and FedProx train collaboratively.The study also compares FedSwitch with fixed switching times to assess its dynamic switching policy.Source: §VI · pp. 7–8
  • The default scenario uses 5 BSs, 7 RoOs, randomly placed BSs, and 112 training samples per BS.Each BS has 16 samples per class across seven regions; the caption reports an average BS distance of 562 m.Source: §VI · p. 8, Figure 2 · p. 8

A. Channel Model

The evaluation uses a 3GPP-based multipath channel model and ResNet-18 classifier, then tests classification performance across system conditions and model compression levels.

A. Channel Model

  • The channel model uses 3GPP-based clustered frequency-selective propagation, with multiple paths and random angle-of-arrival offsets.The BS and transmitter positions determine their distance, and path angles vary around the mean angle of arrival.Source: A. Channel Model · p. 8

B. Network Architecture

  • The classifier uses a ResNet-18 backbone with two input channels for the real and imaginary signal components and a two-layer classification head.The backbone produces a 512-dimensional representation, followed by a 256-unit layer, ReLU, dropout, and an output layer for N classes.Source: B. Network Architecture · pp. 8–9

C. Classification Performance

  • Across tested SNRs, BS distances, and sample counts, FedSwitch is best or nearly best; dynamic switching around round 15 yields the lowest reported error.As BS distance increases, Global, FedAvg, and FedProx degrade; at low sample counts, Single performs poorly, while it becomes more competitive as samples increase.Source: C. Classification Performance · p. 9, Figure 3 · p. 10, Figure 4 · p. 10
  • With compression factor Rk = 0.3, FedSwitch learns slightly faster, but its converged maximum performance is lower than with Rk = 1.Top-k sparsification transmits only the largest-magnitude weights, while the server averages the sparsified models.Source: C. Classification Performance · pp. 9–10, Figure 5 · p. 10

D. Security Performance

FedSwitch performs best in the evaluated security settings, and the Gaussian simulations support the analytical bound linking distribution mismatch to detection performance.

D. Security Performance

  • At BS distance 563 m, FedSwitch achieves the best reported detection trade-off; at 220 m, the gap between Single and aggregated approaches reverses, while FedSwitch remains best.The closer-BS setting makes observed data more i.i.d., improving the effectiveness of information sharing.Source: D. Security Performance · p. 10, Figure 6 · p. 11
  • Across simulated target false-alarm probabilities, empirical missed-detection probability stays below the analytical bound, and greater distribution mismatch worsens performance.The verification uses synthetic Gaussian data with differing legitimate-distribution variances.Source: D. Security Performance · pp. 10–11

VII. CONCLUSIONS

  • The conclusion reports that FedSwitch's loss-monitored local fine-tuning mitigates non-IID challenges and outperforms federated and conventional non-federated baselines in error probability.The evaluations use realistic, 3GPP-compliant channel models.Source: §VII · p. 11

APPENDIX A PROOF OF THEOREM 5.1

The appendix proves the convergence result by bounding the stochastic momentum updates under the theorem's conditions and applying a standard mSGD analysis.

APPENDIX A PROOF OF THEOREM 5.1

  • The proof unfolds the model update at communication steps and bounds the quantity C using Lemma 1.The lemma is stated under the conditions of Theorem 5.1.Source: Appendix A · p. 12
  • The proof applies a standard mSGD convergence result after choosing the stated learning-rate schedule and parameters.It identifies the resulting bound with the form in [19].Source: Appendix A · p. 12
  • Lemma 1 bounds the momentum update by induction, starting from a zero initial update and bounding its expected value at the first step.The proof uses the bound on gradient magnitudes from Assumption 4.Source: Appendix A · p. 12

APPENDIX B PROOF OF THEOREM 5.2

The proof identifies FedAvg with the optimal likelihood-ratio test on the global data distribution, then bounds the per-base-station miss-detection gap using discrepancies between local and global distributions.

APPENDIX B PROOF OF THEOREM 5.2

  • Under the assumptions in Sec. V, FedAvg achieves the optimal LLR test over the global data distribution.The global model is trained on the union of the base stations’ data.Source: Appendix B · p. 12
  • For a given false-alarm probability, the passage defines miss-detection probability and contrasts the global model’s null acceptance region with the Neyman–Pearson test, which is optimal at each base station.The derivation focuses on BS 1 and states that it is symmetric for BS 2.Source: Appendix B · p. 12
  • The decision-region disagreement is bounded through a magnitude inequality and related to total-variation distances between local and mixture distributions.The proof uses equal-weight mixtures of the two base stations’ data distributions.Source: §1 · p. 13, §2 · p. 13, §3 · p. 13
  • Combining the two threshold cases yields the final performance bound in (33).The proof uses reciprocal thresholds for the federated and Neyman–Pearson tests.Source: §4 · p. 13, §1 · p. 13, §3 · p. 13

Related papers

Loading 2610.03523v1…