Gossip-GAN: Decentralized CSI Feedback

• The paper introduces a decentralized CSI-feedback approach where each UE locally trains GANs and exchanges parameters via gossip learning to synthesize CSI without raw data sharing.
• It uses a distributed gossip learning mechanism that reduces uplink bandwidth usage by over 20× while addressing privacy and mobility challenges.
• The method mitigates catastrophic forgetting and achieves NMSE performance close to full retraining, demonstrating robustness across varying channel conditions.

Searching arXiv for the specified paper and closely related CSI-feedback/GAN context. Gossiping Generative Adversarial Network, or Gossip-GAN, is a distributed CSI-feedback training framework for FDD mMIMO-OFDM systems in which each user equipment (UE) trains a local GAN on a small CSI dataset and then exchanges model parameters through gossip learning rather than uploading raw CSI to a base station (BS) (Cao et al., 31 Aug 2025). In the reported formulation, the trained generator is ultimately used at the BS to synthesize a large fake-CSI dataset for training a deep autoencoder (DAE) for CSI feedback. The framework is introduced to address three coupled difficulties of prior DAE-based CSI feedback methods: reliance on large-scale BS-collected data, the resulting bandwidth usage and privacy issues, and degradation under mobility and environment shifts, including catastrophic forgetting when models are retrained across scenarios (Cao et al., 31 Aug 2025).

1. Problem setting and motivating context

The underlying application is CSI feedback in FDD massive multiple-input multiple-output orthogonal frequency-division multiplexing systems. Prior DAE approaches are described as efficient in reducing CSI feedback overhead, but they are also described as relying heavily on large-scale data collected through the BS for model training. In the reported setting, this dependency creates excessive bandwidth usage and data privacy issues, particularly for mMIMO systems (Cao et al., 31 Aug 2025).

The framework is additionally motivated by nonstationary propagation conditions. When users are mobile and encounter new channel environments, existing CSI feedback models may need retraining. Returning to previous environments can then make those models perform poorly and expose them to catastrophic forgetting. Gossip-GAN is proposed specifically to solve these problems by combining local generative modeling, decentralized model exchange, and synthetic-data-based autoencoder training (Cao et al., 31 Aug 2025).

A central property of the proposal is that it is fully decentralized during gossip learning: no raw CSI ever leaves a UE, and only model weights are shared peer-to-peer. This suggests a shift from dataset aggregation to parameter exchange as the primary mechanism for building a usable CSI-data surrogate for downstream feedback training.

2. Signal model and DAE-based CSI feedback formulation

The system model considers a BS with $N_t$ transmit antennas, with $N_t \gg 1$, serving multiple single-antenna UEs over OFDM with $N_c$ subcarriers (Cao et al., 31 Aug 2025). On subcarrier $n$, the received signal is

$$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$

Using the DeepMIMO physical channel model with $L$ paths, the channel vector on subcarrier $n$ is

$$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$

where $\rho_l$, $\vartheta_l$, and $N_t \gg 1$0 denote path power, phase, and delay, and $N_t \gg 1$1 is the 3-D array response given as the Kronecker product of the $N_t \gg 1$2, $N_t \gg 1$3, and $N_t \gg 1$4 steering vectors. The full CSI matrix at a UE is

$N_t \gg 1$5

CSI feedback is implemented through a DAE. The UE first converts $N_t \gg 1$6 into a real tensor of shape $N_t \gg 1$7, then compresses it as

$N_t \gg 1$8

At the BS, reconstruction is performed through

$N_t \gg 1$9

The training objective for the autoencoder is the sample-average MSE,

$N_c$0

Within Gossip-GAN, the DAE is not trained directly from centrally uploaded raw CSI. Instead, the BS trains the AE using synthetic CSI produced by a generator learned in a decentralized manner (Cao et al., 31 Aug 2025).

3. GAN architecture and adversarial objective

The architecture uses a generator $N_c$1 and discriminator $N_c$2 tailored to CSI tensors (Cao et al., 31 Aug 2025). The generator takes as input $N_c$3 of shape $N_c$4. It then applies a ConvTranspose2d layer from 128 to 64 channels with kernel $N_c$5, followed by three ResidualBlock$N_c$6 modules. Each such block upsamples by a factor of 2 via nearest-neighbor interpolation and contains two Conv2d+BN+ReLU layers with 64 channels and a skip-add connection. After a further BN + ReLU stage, a final Conv2d layer maps 64 channels to 2 channels with kernel $N_c$7, padding 1, and tanh activation.

The discriminator receives a real or fake CSI tensor $N_c$8. It contains two ResidualBlock$N_c$9 modules; each includes two Conv2d+BN+ReLU layers with 64 channels, with average-pool downsampling on the first convolution output and on the skip path. Dropout with probability $n$0 is applied after each block. A further ResidualBlock without downsampling is followed by another Dropout stage, after which global spatial averaging yields a 64-dimensional vector that is passed to a Linear layer to produce a scalar logit (Cao et al., 31 Aug 2025).

The losses combine a Wasserstein formulation with a consistency term. The generator objective is

$n$1

The discriminator objective is the simplified CTGAN form

$n$2

where $n$3, $n$4 and $n$5 are two dropout-mask evaluations of $n$6, and $n$7, $n$8, and $n$9 are hyper-parameters. In the reported ablation, the simplified CTGAN reduces compute-graph size by $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$0 with less than $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$1 dB NMSE loss (Cao et al., 31 Aug 2025). This suggests that the consistency regularization was retained in a form judged sufficient for CSI synthesis while reducing computational burden.

4. Gossip-learning mechanism

The defining feature of Gossip-GAN is its distributed gossip-learning strategy. Each UE $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$2 collects local CSI $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$3, initializes $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$4, repeatedly trains the local GAN for $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$5 steps, and then selects a peer $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$6 to which it sends $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$7. On receiving $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$8 from a peer, the UE stores the model. Once $$y_n = \mathbf h_n^H\,\mathbf v_n\,x_n + z_n, \qquad \mathbf h_n\in\mathbb C^{N_t\times1}.$$9 models have been stored, local and received models are averaged as (Cao et al., 31 Aug 2025)

$L$0

The corresponding round-based update for a gossip interaction between UEs $L$1 and $L$2 is

$L$3

where $L$4 denotes the stacked parameters of $L$5, and $L$6 for averaging. Two topologies are reported. In Topology 1, each UE randomly selects 4 neighbors. In Topology 2, the system is fully connected and each UE exchanges with all $L$7 peers (Cao et al., 31 Aug 2025).

Operationally, each UE trains on only a small amount of local data, while gossip exchange is used both to avoid overfitting and to accelerate model training. One trained generator is then sent to the BS, which uses it to synthesize a large fake-CSI dataset for training the AE. A plausible implication is that the global CSI structure is approximated indirectly through repeated peer-to-peer parameter mixing rather than through explicit global data pooling.

5. Privacy, bandwidth overhead, and computational profile

The privacy model is explicit: no raw CSI ever leaves a UE, and only model weights, specifically generator and discriminator parameters, are shared peer-to-peer (Cao et al., 31 Aug 2025). This is the basis for the claim that Gossip-GAN preserves users’ privacy while enabling CSI feedback training.

The reported overhead comparison is numerical and direct. In centralized training, the BS must collect $L$8 real values. For $L$9, $n$0, and $n$1,

$n$2

Under Gossip-GAN, each UE uploads only its generator at the end of gossip learning. The generator contains approximately $n$3 M weights, corresponding to

$n$4

which is reported as a $n$5 reduction in uplink (Cao et al., 31 Aug 2025). The paper further states that fully decentralized gossip learning reduces uplink overhead by $n$6 versus centralized dataset upload.

The per-epoch computational profile on the DeepMIMO sparse scenario is also quantified. Gossip-GAN requires $n$7 T FLOPs, $n$8 MB memory, and $n$9 s/epoch, whereas centralized GAN training requires $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$0 T FLOPs, $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$1 MB memory, and $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$2 s/epoch (Cao et al., 31 Aug 2025). The computational complexity per UE is given as

$$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$3

These figures place Gossip-GAN not only as a communication-efficient method but also as a lighter-weight training strategy at the UE level.

6. Continual adaptation, robustness, and empirical performance

A major use case is continual adaptation across channel scenarios. The reported failure mode is that a CSI feedback DNN trained in one area, such as a sparse area, fails in another, such as a dense area, and vice versa (Cao et al., 31 Aug 2025). To address this, the continual-learning variant stores the trained generator $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$4 for each encountered scenario $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$5. When the system enters a new scenario $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$6, it forms the mixed synthetic dataset

$$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$7

The AE is then retrained on $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$8. No real CSI is stored; only small generator files of approximately $$\mathbf h_n = \sum_{l=1}^L \sqrt{\rho_l\,N_c}\, e^{j\bigl(\vartheta_l + \tfrac{2\pi n}{N_c} \tau_l B\bigr)} \,\mathbf a(\phi_{az}^l,\phi_{el}^l),$$9 M parameters each are retained (Cao et al., 31 Aug 2025).

Catastrophic forgetting is evaluated through the drop in NMSE on an old scenario after adaptation to a new one. In the sparse-to-dense transition example, no retraining yields an NMSE of $\rho_l$0 dB in the sparse area, full retraining with stored real CSI recovers to $\rho_l$1 dB, and Gossip-GAN continual learning using stored generators achieves $\rho_l$2 dB in the sparse area and $\rho_l$3 dB in the dense area (Cao et al., 31 Aug 2025). The reported summary is that Gossip-GAN continual learning recovers old-scene NMSE to within 1 dB of full retraining while using only generator storage.

The simulation setup uses the DeepMIMO “O1_28” outdoor 28 GHz scenario. The dense area corresponds to rows 3540–3600 sampled at $\rho_l$4 m, and the sparse area to rows 1000–1400 sampled at $\rho_l$5 m. The bandwidth is $\rho_l$6 GHz, with $\rho_l$7, $\rho_l$8, and $\rho_l$9. A total of $\vartheta_l$0 samples are split into $\vartheta_l$1 training and $\vartheta_l$2 test examples and normalized to $\vartheta_l$3. The AE is trained with learning rate $\vartheta_l$4, 100 epochs, and $\vartheta_l$5. The GAN is trained with learning rate $\vartheta_l$6, 1000 epochs, $\vartheta_l$7, $\vartheta_l$8, $\vartheta_l$9, dropout $N_t \gg 1$00, and batch size 100 (Cao et al., 31 Aug 2025).

Performance is measured by NMSE,

$N_t \gg 1$01

along with convergence speed, computational cost, and bandwidth reduction. The reported empirical findings include the following. A “No connection” GAN trained with 500 CSI samples per UE and no gossip yields approximately $N_t \gg 1$02 dB in the dense area and approximately $N_t \gg 1$03 dB in the sparse area. Gossip-GAN Topology 2 is approximately equal to centralized GAN, with $N_t \gg 1$04 dB in the dense area and $N_t \gg 1$05 dB in the sparse area. Federated GAN (SyncDG) is slightly better, but it incurs 1.6 M parameters per BS round. NMSE improves sharply up to $N_t \gg 1$06 fake samples and then saturates. Across compression ratios $N_t \gg 1$07, Gossip-GAN remains within 0.5 dB of centralized training. Increasing the number of UEs from $N_t \gg 1$08 to $N_t \gg 1$09 improves NMSE by 1–2 dB. Compatibility is also reported with AE back-ends CsiNet, CRNet, CsiQNet, and DeepCMC, with improvements of 0.2–0.5 dB for each (Cao et al., 31 Aug 2025).

The abstract additionally reports that the proposed approach possesses an inherent robustness. In context, this appears alongside robustness to environment changes and the synthetic-replay mechanism used against forgetting.

7. Limitations and prospective extensions

The reported limitations are specific. There is a slight performance gap versus Federated GAN with a server-aided design, which motivates possible hybrid topologies. Additional future directions listed are adversarial-ML defenses for the GAN pipeline, exploration of richer generative models such as diffusion and normalizing flows, application to beamforming, channel estimation, and data detection in the physical layer, and validation under 3GPP TR 38.901 channel models and highly non-iid UE distributions (Cao et al., 31 Aug 2025).

These points delimit the present scope of Gossip-GAN. The method is positioned as a fully decentralized alternative for low-overhead CSI feedback training, with peer-to-peer model averaging, synthetic CSI generation, privacy preservation through non-disclosure of raw CSI, and generator-based replay for catastrophic-forgetting mitigation. The reported results indicate near-centralized CSI-model quality under decentralized training, but they also indicate that server-aided federated variants can still provide slightly better performance under the evaluated conditions (Cao et al., 31 Aug 2025).