PaperlayerView exact revision on arXiv

Source-linked AI summary

FedGenSC: Federated Generative Semantic Communication with Channel-Aware Adaptation

Rita Abou Fares, Razan Al Kakoun, Maher Nouiehed, Hadi Sarieddeen

arXiv:2609.08593v1eess.SPcs.LG

TL;DR

Naive GAN federation in semantic communication raises unresolved challenges involving discriminator aggregation, divergent client embedding spaces, and generation that does not adapt to channel conditions. FedGenSC combines a global generator with local-only discriminators, a semantic prototype bank, and SNR-conditioned generation. On Europarl over Rayleigh fading, it improves BLEU-1 over the non-IID FedDeepSC baseline by up to 58.2%.

  • Problem

    Federated GAN-based SemCom faces three challenges: discriminator aggregation instability under non-IID data, semantic drift from divergent local embeddings, and channel-agnostic generation under heterogeneous link conditions.

    Source: §I · p. 1

  • Method

    FedGenSC federates a generator while keeping discriminators local, uses a semantic prototype bank for cross-client grounding, and conditions generation on channel state.

    Source: §I · p. 2

  • Results

    FedGenSC achieves up to 58.2% BLEU-1 improvement over the non-IID FedDeepSC baseline on Europarl under Rayleigh fading.

    Source: §V · p. 6

  • Takeaways & Limitations

    The results highlight the potential of generative, channel-aware federated SemCom for future wireless systems.

    Source: §V · p. 6

Abstract

from arXiv · show

Integrating generative adversarial networks (GANs) into federated semantic communication (SemCom) is a natural progression, as generative priors can recover semantic fidelity under channel distortion that discriminative decoders cannot. However, naive GAN federation introduces three failure modes that prior work has, to the best of our knowledge, neither identified nor resolved: discriminator aggregation instability under non-independent and identically distributed (non-IID) data, semantic drift caused by divergent local embedding spaces, and channel-agnostic generation that cannot adapt to heterogeneous link conditions. We propose federated generative semantic communication (FedGenSC), which mitigates all three by employing a global generator with local-only discriminators, providing cross-client semantic information through a semantic prototype bank, and conditioning generation on the instantaneous signal-to-noise ratio (SNR). Experiments on the Europarl dataset over Rayleigh fading channels (K=10 clients, Dirichlet {\alpha}=0.5) show that FedGenSC under non-IID data outperforms the FedDeepSC baseline across the tested SNR range, achieving up to a 58.2% relative improvement in bilingual evaluation understudy (BLEU)-1 at 18 dB. Ablation studies confirm the independent contribution of each component.

I. INTRODUCTION

The introduction motivates federated generative semantic communication as a response to three unresolved challenges in GAN-based federated SemCom. FedGenSC uses a shared generator, local discriminators, semantic prototypes, and SNR conditioning; the system model specifies distributed text data and fading channels.

I. INTRODUCTION

  • Prior federated SemCom work uses discriminative decoders, while federated GAN deployment raises unresolved problems of discriminator instability, semantic drift, and channel-agnostic generation.The introduction identifies these three challenges as not jointly characterized and addressed in prior federated SemCom literature.Source: §I · p. 1
  • FedGenSC federates only the generator, keeps discriminators local, uses a semantic prototype bank, and conditions generation on instantaneous SNR.It performs semantic-space pre-equalization using a scalar SNR estimate.Source: §I · pp. 1–2

A. System Model

  • The system model has a central server and clients with local text datasets drawn from potentially different distributions, each using a local DeepSC semantic codec.The encoder produces token embeddings, the channel encoder compresses them, and the receiver reconstructs the sentence.Source: A. System Model · p. 2
  • Clients communicate over independent Rayleigh fading channels with client-specific instantaneous SNR and additive white Gaussian noise.The transmitted signal is normalized to unit average power.Source: A. System Model · p. 2

B. Federated Training Objective

The federated objective minimizes expected semantic distortion across clients while preserving the constraint that raw data are not shared.

B. Federated Training Objective

  • The shared parameters comprise the semantic encoder, channel encoder, decoder, and generator, and are optimized to minimize expected semantic distortion across clients.The objective is constrained by the requirement that clients do not share raw data.Source: B. Federated Training Objective · p. 2
  • The semantic distortion is token-level cross-entropy, while BLEU-1 evaluates reconstruction quality.Because raw data remain private, centralized DeepSC is not a suitable comparison under this constraint.Source: B. Federated Training Objective · p. 2
  • The no-raw-data-sharing constraint makes centralized DeepSC unsuitable as a comparison because it requires pooling clients’ datasets on one server.Source: B. Federated Training Objective · p. 2

C. Challenges of Integrating GANs into Federated SemCom

The section describes discriminator aggregation instability and semantic drift as two challenges in federated GAN-based SemCom. It relates each to non-IID client data and reports empirical degradation or improvement associated with addressing the challenge.

1) P1 — Discriminator Aggregation Instability:

  • Discriminator aggregation degrades BLEU-1 by 61.8% at 6 dB and falls below the no-GAN baseline at mid-to-high SNR.Under non-IID data, local discriminators can develop incompatible decision boundaries, so their averaged boundary may not match any client’s local minimax solution.Source: 1) P1 — Discriminator Aggregation… · p. 2

2) P2 — Semantic Drift:

  • Divergent local embedding spaces can make the aggregated generator produce embeddings that deviate from client representations, a phenomenon termed semantic drift.The same word or phrase can map to different regions across clients; addressing the problem yields a reported 62.8% BLEU-1 improvement at 6 dB.Source: 2) P2 — Semantic Drift · p. 2

3) P3 — Channel-Agnostic Generation:

Channel-agnostic generation cannot tailor corrections to clients’ heterogeneous instantaneous SNRs. The proposed semantic-space generator takes SNR and a prototype alongside the sentence embedding, then produces a residual correction before channel encoding.

3) P3 — Channel-Agnostic Generation:

  • Clients’ heterogeneous instantaneous SNRs make a channel-agnostic generator unable to tailor its output to channel conditions.A channel-agnostic generator neglects dependence on SNR.Source: 3) P3 — Channel-Agnostic Generation · p. 2
  • SNR conditioning reduces BLEU-1 degradation by up to 49.4% at 6 dB in the controlled ablation.Without channel dependence, generation can under-correct at low SNR or over-correct at high SNR.Source: 3) P3 — Channel-Agnostic Generation · p. 2

A. Overview and Design Principles

  • The conditional generator takes the mean-pooled sentence embedding, instantaneous SNR, and semantic prototype, and produces a residual correction added before channel encoding.Source: A. Overview and Design Principles · pp. 2–3
  • The discriminator compares the original and generator-corrected embeddings before channel encoding, training corrections to resemble clean embeddings in the semantic space.Working in the lower-dimensional semantic space reduces the complexity of learning meaningful corrections relative to raw signal space.Source: A. Overview and Design Principles · p. 3

B. System Architecture

FedGenSC inserts a conditional GAN between DeepSC’s semantic and channel encoders. The generator applies an SNR-conditioned residual correction to semantic embeddings, while a local discriminator evaluates corrected embeddings before transmission.

B. System Architecture

  • The Transformer encoder maps tokenized input to a sequence of semantic embeddings E with model dimension d=128.The pipeline places its conditional GAN between DeepSC’s semantic and channel encoders.Source: B. System Architecture · p. 3
  • The generator produces a residual correction that is added to the encoder output and broadcast across all T token positions.Its input includes the sentence embedding and instantaneous SNR, as indicated in the supplied method passages.Source: B. System Architecture · p. 3
  • A local discriminator evaluates the pair of original and corrected embeddings; the corrected sequence then passes to the channel encoder for transmission.The discriminator is instantiated locally and discarded after local training.Source: B. System Architecture · p. 3

C. Semantic Prototype Bank

The semantic prototype bank summarizes client embedding spaces with transmitted centroids, then supplies global semantic anchors. Local training alternates DeepSC fine-tuning with GAN updates while keeping discriminators local.

C. Semantic Prototype Bank

  • Each client summarizes its local semantic space with k-means centroids and transmits only those vectors, avoiding raw-data sharing.The server combines local summaries into a global bank that is redistributed to clients; the passages do not specify the number of centroids here.Source: C. Semantic Prototype Bank · p. 3
  • For each input embedding, the nearest centroid in the global bank serves as a semantic anchor to counteract misaligned embedding geometries and support robustness across clients.Each batch embedding receives its own prototype; communication is P × d values per client per round, or 2,560 values when P=20 and d=128.Source: C. Semantic Prototype Bank · p. 3

D. Two-Timescale Local Training Protocol

  • Local training first fine-tunes DeepSC for three epochs with semantic loss, then freezes the encoder and trains a fresh discriminator with the generator initialized from the global model.The update scheme uses five discriminator steps per generator step.Source: D. Two-Timescale Local Training Protocol · p. 3
  • The discriminator distinguishes clean reference pairs from generated pairs, while generator training keeps corrections near the original embedding and aligns corrected embeddings with valid sentence representations.Residual regularization limits deviation; the adversarial term encourages manifold alignment.Source: D. Two-Timescale Local Training Protocol · pp. 3–4
  • Prototype alignment informs generator aggregation, and clients transmit their updated DeepSC and generator parameters, representative SNR, and local prototype bank—but not the discriminator.Local prototype banks are periodically aggregated at the server to provide common prototype information.Source: D. Two-Timescale Local Training Protocol · p. 4

E. Federated Generator Aggregation

FedGenSC forms the global generator by weighted averaging rather than treating all clients as equally reliable. The weights reflect channel representativeness, generator quality, and semantic alignment.

E. Federated Generator Aggregation

  • The global generator is computed as a weighted average of client generators, with weights representing client reliability.Naive averaging can be suboptimal when clients differ in SNR, generator quality, or semantic alignment.Source: E. Federated Generator Aggregation · p. 4
  • The aggregation weights combine three quality signals: channel representativeness, generator accuracy, and alignment between local and global semantic structures.The supplied passages identify these signals as SNR, generator, and prototype quality.Source: E. Federated Generator Aggregation · p. 4
  • Combining the signals is designed to reduce outlier-client influence and limit overfitting to one SNR regime or local data distribution.Local discriminator parameters are neither transmitted nor aggregated, decoupling discriminator training across clients.Source: E. Federated Generator Aggregation · p. 4

A. Experimental Setup

Experiments evaluate FedGenSC on Europarl over Rayleigh fading with ten non-IID clients and a range of SNRs. FedDeepSC removes the GAN module, while centralized DeepSC is excluded under the no-raw-data-sharing constraint.

A. Experimental Setup

  • Experiments use Europarl with K=10 clients partitioned by a Dirichlet distribution with α=0.5, over Rayleigh fading at SNRs from 0 to 18 dB.Evaluation spans 50 rounds and reports BLEU-1 and SBERT cosine similarity; each client uses 20 prototype centroids.Source: A. Experimental Setup · p. 4
  • FedDeepSC baselines match FedGenSC without the GAN, though some observed gain may reflect added GAN capacity; centralized DeepSC is excluded because it requires pooling client data.The exclusion follows the study’s no-raw-data-sharing constraint.Source: A. Experimental Setup · p. 4

B. Main Experiments

FedGenSC improves BLEU-1 over federated baselines across the tested SNR conditions, with gains also reflected in SBERT similarity. Ablations report substantial losses when any of the three design components is removed.

a) Non-IID setting

  • 58.2% relative BLEU-1 improvement over the non-IID baseline at 18 dB, with FedGenSC exceeding both non-IID and IID baselines across the full SNR range.The non-IID evaluation uses Rayleigh fading with Dirichlet α=0.5 and K=10 clients.Source: B. Main Experiments · p. 5, Table II · p. 5

b) IID setting

  • FedGenSC consistently improves over the IID baseline, with larger gains at low-to-mid SNR where GAN correction is most beneficial.At high SNR, the prototype-bank benefit diminishes as channel-induced distortions become smaller.Source: B. Main Experiments · p. 5

Main results

  • BLEU-1 gains are accompanied by higher SBERT similarity, indicating improved semantic fidelity beyond direct word overlap.Table IV reports both BLEU-1 and SBERT scores across SNR.Source: B. Main Experiments · p. 5, Table IV · p. 5

Ablation results

  • At 6 dB, naive GAN federation drops 61.8%, prototype removal drops 62.8%, and SNR-conditioning removal drops 49.4% versus full FedGenSC.The ablation table covers BLEU-1 across the full SNR range under non-IID partitioning.Source: B. Main Experiments · p. 5, Table III · p. 5

C. Ablation Study

FedGenSC combines a shared generator, local discriminators, a semantic prototype bank, and SNR conditioning. Its non-IID results and component-removal tests report gains over baselines and independent contributions from all three components.

C. Ablation Study

  • Naive GAN federation, which aggregates the discriminator, causes the most severe degradation (−61.8% at 6 dB), yet exceeds the no-GAN baseline at 6, 9, 12, and 15 dB and falls below it only at 18 dB.Source: C. Ablation Study · p. 5
  • Removing prototype conditioning causes the largest single-component drop at 6 dB; without a global semantic reference, client generators may specialize locally and the aggregated generator may drift toward a compromise serving no client well.Source: C. Ablation Study · p. 5
  • Removing SNR conditioning produces the smallest drop among the three ablations, and this drop shrinks further as SNR increases; at low SNR, precise channel conditioning is likely important.Source: C. Ablation Study · p. 5
  • All three components contribute independently across the full SNR range in the evaluated setting.Prototype removal and naive GAN federation cause the largest drops; naive GAN federation also falls below the no-GAN baseline at high SNR.Source: C. Ablation Study · p. 5

V. CONCLUSION

  • Under IID partitioning, FedGenSC outperforms FedDeepSC across the full SNR range, with larger gains at lower SNR.The figure covers Rayleigh fading with K=10 clients.Source: Figure 1 · p. 6
  • FedGenSC achieves up to 58.2% BLEU-1 improvement over the non-IID baseline on Europarl under Rayleigh fading.It exceeds the non-IID baseline across the tested SNR range with Dirichlet α=0.5 and K=10 clients.Source: §V · p. 6, Figure 2 · p. 6
Loading 2609.08593v1…