{"id":"bbf8cbfc-85ce-4b30-b9b3-91310a957d64","arxiv_id":"2509.10490","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Gossip-GAN trains a distributed GAN across user devices and uses its generated channel samples to train a CSI feedback autoencoder, matching centralized training accuracy while reducing uplink overhead.","lead":"This paper trains GAN models on user devices, using gossip-style communication between devices to combine their models without uploading channel data to the base station. The resulting generator creates synthetic channel data used to train a CSI feedback autoencoder, achieving accuracy near centralized training while reducing uplink bandwidth and helping avoid catastrophic forgetting when users move.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniform weight averaging in Eqs. (15)-(16) is the load-bearing step; it is supported only in an IID-shard experiment, not in the non-IID regimes the distributed framework targets.","rationale":"The reader identified the model-averaging step in Eqs. (15)-(16) as the weakest assumption, and I agree that this is the most load-bearing point. The central claim requires that a uniformly averaged GAN generator behaves like a GAN trained on the aggregate CSI distribution. The paper provides indirect evidence through Fig. 8 for IID shards, which is real but incomplete evidence, and it provides no theoretical justification. Because the framework is explicitly distributed and motivated by heterogeneous real deployments, the missing non-IID validation is a correctness risk rather than merely a stylistic gap. I am not moving the verdict from CONDITIONAL: the concern is concrete and testable, but it does not by itself contradict the paper's reported results; it limits their generality. The reader's CONDITIONAL verdict already captures this, so my read leaves the verdict unchanged. I also flag the stale-buffer bug in Algorithm 1 as a secondary implementation inconsistency, but the non-IID averaging test is the single concrete check that would settle whether the central claim holds beyond the current setup.","tokens_in":26056,"tokens_out":8612,"duration_ms":106039,"concrete_test":"Run the proposed Gossip-GAN (topology 2, K=10, 500 samples/UE, same hyperparameters as Section IV-B) on deliberately non-IID partitions: draw each UE's 500 samples from a different convex combination of the sparse and dense DeepMIMO sets, keeping the total at 5000. Then compare (a) the FID/MMD between the merged generator's output and the true aggregate CSI distribution, and (b) the downstream CsiNet NMSE, against the centralized-GAN baseline trained on the same 5000 samples. If the non-IID NMSE gap exceeds the IID gap by more than about 1 dB, or if the distributional metric degrades sharply, Eqs. (15)-(16) are not a valid aggregation rule for non-IID data, and the central claim must be restricted to the IID-shard setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire Gossip-GAN pipeline depends on Eqs. (15)-(16) of Section III-B: averaging generator and discriminator weights across gossip peers should produce a GAN whose generator approximates the aggregate CSI distribution. This is not a harmless approximation. GAN training is a non-convex min-max problem; averaging two trained WGAN/CTGAN models can produce a generator with collapsed modes or inconsistent BatchNorm statistics. The paper cites only a position paper [46] for this rule and offers no convergence analysis, no aggregation-error bound, and no direct distributional check of the merged generator.\n\nThe empirical support in Section IV-B is also narrower than the claim. Each of K=10 UEs receives 500 samples from the same dense-area or sparse-area dataset, so the local shards are approximately IID. Real mMIMO deployments motivated by the framework have UEs at different locations, BS sectors, or mobility conditions, so local CSI distributions differ. No experiment varies the degree of non-IID-ness across UEs, and this is exactly where uniform averaging of GAN parameters is known to be fragile. If Eqs. (15)-(16) fail, the synthetic dataset, autoencoder training, and catastrophic-forgetting results all collapse.\n\nA secondary internal issue: Algorithm 1 never clears the saved received-models buffer after a merge, so a literal implementation would keep stale models and include them in all subsequent merges. This should be corrected or clarified before the merging rule can be assessed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Gossip-GAN, a fully distributed framework for training CSI feedback autoencoders in FDD mMIMO-OFDM systems. Each UE collects a small local CSI dataset and trains a local GAN (a WGAN-GP/CTGAN variant), then gossips generator/discriminator weights with neighboring UEs; the averaged generator is sent to the BS, which synthesizes a large fake CSI dataset used to train a CsiNet-type autoencoder. The paper also applies the framework to catastrophic forgetting by storing past-scenario generators and replaying their samples when training on new scenarios. Experiments on DeepMIMO and COST2100 datasets compare NMSE against centralized GAN training, no-connection distributed GAN, federated learning, and true CSI training, and report comparable or slightly worse NMSE with lower uplink overhead and memory cost.","tokens_in":26458,"tokens_out":6418,"duration_ms":79899,"significance":"If the central claims hold, the framework is a meaningful step toward privacy-preserving, low-overhead distributed training for CSI feedback, and it demonstrates compatibility with several modern autoencoder architectures (CsiNet, CRNet, CsiQNet, DeepCMC). The paper includes useful empirical elements: public DeepMIMO data, repeated experiments with 95% confidence intervals, a comparison against federated GAN, and a memory-cost comparison for catastrophic forgetting. However, the load-bearing model-averaging step (Eqs. 15–16) is supported only by an IID-shard experiment and no convergence or distributional analysis, so the significance is currently conditional on additional validation in heterogeneous settings.","major_comments":[{"comment":"Uniform averaging of generator and discriminator weights across gossip peers is the load-bearing step of the proposed framework, but it is supported only by a citation to a position paper [46]. GAN training is a non-convex min-max problem, and averaging two trained WGAN/CTGAN models is not generally guaranteed to produce a generator that approximates the aggregate CSI distribution; mode collapse or inconsistent BatchNorm statistics are real risks. The paper provides no convergence analysis, no aggregation-error bound, and no direct distributional check (e.g., MMD, FID, or generated-sample NMSE before autoencoder training) of the merged generator. I request either theoretical justification or a direct empirical validation of the merged model, especially before the downstream autoencoder training is shown to depend on it.","section":"§III-B, Eqs. (15)–(16)"},{"comment":"The experimental setup uses K=10 UEs, each receiving 500 samples randomly drawn from the same dense-area or sparse-area dataset. This makes the local shards approximately IID, which is the most favorable regime for uniform parameter averaging. The framework is motivated by UEs at different locations, BS sectors, or mobility conditions, where local CSI distributions differ. No experiment varies the degree of non-IID-ness across UEs. Since uniform averaging is known to degrade under heterogeneity, the claim that Gossip-GAN 'can achieve similar CSI feedback accuracy as centralized training' is not yet established for the target deployment scenarios. Please add experiments with controlled non-IID shards (e.g., UEs assigned disjoint row ranges of DeepMIMO or different BSs) and report per-UE distribution shift and final NMSE.","section":"§IV-B, Fig. 8"},{"comment":"In the ONRECEIVEMODEL function, received models are saved to a buffer, and merging is triggered when the buffer size reaches npeers. The buffer is never cleared after MERGE_SAVED_MODELS, so a literal implementation would include stale models in every subsequent merge, rather than averaging the most recent npeers peer models. This is an algorithmic bug or an under-specification that affects the correctness and reproducibility of the gossip procedure. Please correct the pseudo-code (e.g., clear the buffer after merging) and also define npeers explicitly for both topologies (for Topology 2, presumably npeers=K−1) and specify the communication interval Δ used in the simulations.","section":"Algorithm 1, lines 12–16"},{"comment":"The number of synthetic samples S is set to 1.0e4 after observing that the test NMSE saturates at that point in Fig. 10. Selecting a hyperparameter based on the test-set performance curve is a form of test-set leakage and makes the reported NMSE optimistic. S should be chosen on a validation split (or justified by a separate model-selection procedure), and the sensitivity of the final NMSE to S should be reported. This is load-bearing because the claim of 'similar accuracy to centralized training' depends on the chosen S; without a clean selection protocol, the reader cannot assess whether the reported numbers are tuned to the test set.","section":"§IV-B, Fig. 10 and S=1e4"}],"minor_comments":[{"comment":"The abstract states that results are obtained with 'real-world datasets,' but DeepMIMO is a ray-tracing simulation dataset and COST2100 is a stochastic channel-model dataset. Please rephrase to 'simulation-based channel datasets' or clarify what is meant by real-world.","section":"Abstract and §IV-A"},{"comment":"The introduction states 'Section III provides the simulation results' and 'Section IV concludes our paper,' but the actual structure has simulations in Section IV and conclusions in Section V. Please correct the section references.","section":"§I"},{"comment":"Reference [10] is identical to reference [2]. Please remove the duplicate.","section":"References"},{"comment":"The claim of reduced uplink bandwidth is not quantified against a concrete raw-CSI upload baseline. Please provide a table or calculation showing the number of bits for raw CSI (e.g., 32×32×2 complex values per sample × 5000 samples) versus the transmitted model parameters (0.455M floats), and also account for the D2D gossip traffic that occurs during training.","section":"§IV-B and Table IV"},{"comment":"For Topology 2, the text says 'each UE communicates with all the other UEs simultaneously,' but Algorithm 1 does not define npeers for this case. Please state explicitly that npeers=K−1 for Topology 2 and clarify whether the save/merge buffer is reset after each aggregation round.","section":"§III-B, Fig. 5"},{"comment":"The table layout is confusing: 'After training on' is used as a column header, but the rows mix the training environment and the evaluation environment. Please reformat the table so that rows clearly indicate (training environment, evaluation environment) pairs, and add a sentence explaining the catastrophic-forgetting phenomenon quantitatively.","section":"§IV-C, Table VI"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a timely and relevant problem, but the key novelty—gossip-based GAN averaging—is not sufficiently validated for the non-IID conditions that motivate the work. The absence of heterogeneous-shard experiments and the test-set-based selection of S are the main blockers. The related Dig-CSI framework [22] is closely connected and should be discussed and ideally compared, as it also uses a generative model for distributed CSI feedback training."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Take: this is a workmanlike engineering paper, not a theoretical breakthrough. The new thing is the combination—gossip-trained GANs that synthesize CSI for training a feedback autoencoder, with generative replay to handle catastrophic forgetting. That combination is plausible and, to my knowledge, genuinely first in the CSI feedback literature.\n\nWhat the paper does well: it runs real experiments on DeepMIMO and COST2100, compares against centralized training, non-connected GAN, federated GAN, and retraining, and reports 20-repeat confidence intervals. The results show the framework approximating centralized training in their setups. That is real evidence. The memory comparison against storing true CSI is also useful and fairly presented.\n\nThe soft spot is the load-bearing step: Eqs. (15)-(16) uniformly average generator and discriminator weights across gossip peers. The paper offers no convergence argument, no aggregation-error bound, and no direct distributional check of the merged generator. It cites a position paper for this rule. Worse, the experiments don't stress it: all K=10 UEs receive 500 samples from the same dense or sparse area, so the shards are approximately IID. Real deployments motivated by the framework would have UEs with different locations, sectors, or mobility, i.e., non-IID data. If uniform averaging degrades under non-IID, the entire pipeline collapses. The authors need at least one experiment that varies the degree of non-IID-ness, and ideally a direct check on the aggregate distribution.\n\nThere are also two smaller issues. First, Algorithm 1 never clears the saved-models buffer after a merge, so a literal implementation would include stale models in all subsequent merges. That's a bug or a clarity problem and should be fixed. Second, the choice of S=1e4 synthetic samples is post hoc, based on saturation in Figure 10. It's a reasonable practical choice, but there is no held-out validation, and no baseline like data augmentation with true CSI. The catastrophic-forgetting claims could also be compared against continual-learning baselines (EWC, online learning) rather than only retraining with stored data.\n\nSo: the central idea is plausible, and the paper is a decent engineering contribution. But the load-bearing averaging step is under-validated, and the IID-only experiments are too narrow. I'd send it to review, not desk-reject, but I'd push for additional experiments and a fix of the buffer issue. It's the kind of paper that could be solid after revision.","headline":"A credible engineering combination of gossip learning and GANs for CSI feedback; the parameter-averaging step is the load-bearing assumption and it is not adequately validated.","tokens_in":26895,"tokens_out":4245,"would_cite":true,"duration_ms":51507,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a gossip-trained GAN can substitute for centralized training of CSI-feedback autoencoders in FDD massive MIMO, matching accuracy while cutting uplink overhead.","keywords":["CSI feedback","massive MIMO","generative adversarial network","gossip learning","catastrophic forgetting","deep autoencoder","FDD mMIMO-OFDM"],"falsifier":"Train ten GANs on disjoint local CSI subsets, merge weights by Eq. (15), and compare the merged generator's output distribution against the true mixture distribution using a distribution distance such as Wasserstein distance; if the merged generator is far from the mixture, or if the autoencoder trained on its synthetic data fails to reach the reported NMSE range on held-out channels, the averaging step is the point of failure.","tokens_in":25988,"feed_emoji":"📡","tokens_out":4658,"duration_ms":58726,"temperature":0.7,"pith_summary":"This paper tries to establish that the neural network used to compress channel state information (CSI) in frequency-division duplex massive MIMO systems can be trained without shipping raw channel data to the base station. Its Gossip-GAN scheme has each participating user train a small generative adversarial network on locally collected channels, periodically exchange generator and discriminator weights with a few peers over device-to-device links, and average the received weights into the local model. One finished generator is then sent to the base station, which uses it to synthesize a large dataset for training the compressor-reconstructor autoencoder. The authors report that this scheme matches the feedback accuracy of centralized training on the DeepMIMO and COST2100 datasets, requires about ten times less computation per epoch than a centrally trained GAN, and lets a mobile user revisit an old environment without retraining by storing one small generator per environment instead of raw data.","feed_headline":"Gossip-trained GAN matches centralized CSI feedback accuracy","feed_subtitle":"Users swap only small GAN weights, cutting uplink bandwidth and keeping old environments from being forgotten.","key_machinery":"The load-bearing object is the averaged-weight Gossip-GAN: each user trains a generator and discriminator on local CSI, then every fixed interval sends the weights to a peer selected by a fixed topology, and when enough models arrive merges them by uniform averaging (Eqs. 15-16). The paper relies on this averaged model to be a valid GAN approximating the aggregate channel distribution; it then selects one trained generator, transmits its 0.455M parameters to the base station, and samples a synthetic dataset of size 10^4 to train the CsiNet autoencoder. For catastrophic forgetting, the same generator is stored per environment and the mixed dataset from all stored generators is used for traini","core_discovery":"The central discovery is that a fully distributed gossip-learning GAN, with no central server, can capture the channel distribution well enough that a single user's generator, forwarded to the base station, produces synthetic CSI on which a deep autoencoder achieves feedback accuracy close to training on true centralized data. The authors show this in two DeepMIMO scenarios and with COST2100, find performance improves with more participating users, and demonstrate the anti-forgetting property: after moving from a sparse to a dense area, combining the stored old-environment generator with the new one yields NMSE of -14.75 dB in the old area and -18.51 dB in the new one, versus -0.33 dB withou","pith_inferences":["If uniform weight averaging of GANs is as valid as the paper assumes, the same gossip training recipe should transfer to other generative models such as VAEs or diffusion models for channel synthesis, with the same bandwidth savings.","The generator-per-environment memory strategy is essentially a parameter-based continual-learning mechanism; it could be combined with rehearsal-free techniques like elastic weight consolidation to retain more than two environments without storing a generator per scene.","Because only synthetic data reach the base station, the scheme could be paired with differential privacy on local gradients to blunt the adversarial-machine-learning risks the authors list as future work.","The reported bandwidth advantage grows with antenna count, so the scheme should become more attractive in larger mMIMO arrays; a direct scaling study would quantify that crossover."],"forward_implications":["Training a CSI-feedback autoencoder no longer requires uploading raw channel data to the base station; only the parameters of one generator are transmitted, and that overhead stays constant as antenna arrays scale up.","The distributed gossip strategy cuts per-epoch training computation and memory by roughly a factor of ten versus centralized GAN training, making generative training feasible on resource-constrained users.","Users returning to a previously seen environment keep useful feedback accuracy because each environment is represented by a compact generator, not raw data, and the memory cost drops from 39.06M to 1.82M units versus retraining.","The scheme is architecture-agnostic: replacing CsiNet with CRNet, CsiQNet, or DeepCMC preserves or improves NMSE.","Fully connected gossip topology approximates centralized GAN performance, and accuracy improves as the number of participating users grows from 5 to 10."],"supporting_citations":[{"why":"Supplies the CsiNet autoencoder baseline and the COST2100 dataset used for CSI feedback training.","marker":"[17]"},{"why":"Provides the CTGAN consistency-regularized Wasserstein GAN loss adopted for the local generators.","marker":"[30]"},{"why":"Supplies the gossip-learning model averaging operation (Eqs. 15-16) that merges generator and discriminator weights.","marker":"[46]"},{"why":"Supplies the DeepMIMO ray-tracing dataset and channel model used for the two evaluation scenarios.","marker":"[35]"},{"why":"Provides the CRNet multi-resolution architecture used to test compatibility of the Gossip-GAN training.","marker":"[18]"},{"why":"Provides the federated GAN (Sync D&G) baseline compared for NMSE and communication overhead.","marker":"[48]"},{"why":"Supplies the user-centric gossip training idea for autoencoder CSI feedback that motivates the distributed strategy.","marker":"[27]"},{"why":"Establishes that GAN-generated channels match the real channel distribution, supporting the use of a generator for synthetic CSI.","marker":"[23]"}],"fun_headline_variants":["Gossip-GAN cuts CSI feedback training overhead","Distributed GAN training for CSI without central server","Gossip GAN matches centralized CSI accuracy, cuts bandwidth","Anti-forgetting CSI feedback via gossip GAN","Users swap weights, not data: CSI feedback via gossip GAN"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Everything downstream depends on the assumption that averaging the weights of peer-trained GANs yields a single generator whose synthetic samples still look like real channels from the combined user region; the paper does not prove this and borrows the averaging step from a position paper.","fun_headline_variants_meta":{"raw":{"variants":["Gossip-GAN cuts CSI feedback training overhead","Distributed GAN training for CSI without central server","Gossip GAN matches centralized CSI accuracy, cuts bandwidth","Anti-forgetting CSI feedback via gossip GAN","Users swap weights, not data: CSI feedback via gossip GAN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1203,"prompt_tokens":790,"completion_tokens":413,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":333}},"tokens_in":534,"tokens_out":413,"duration_ms":4498,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:16:38.243484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ten GANs on disjoint local CSI subsets, merge weights by Eq. (15), and compare the merged generator's output distribution against the true mixture distribution using a distribution distance such as Wasserstein distance; if the merged generator is far from the mixture, or if the autoencoder trained on its synthetic data fails to reach the reported NMSE range on held-out channels, the averaging step is the point of failure.","supporting_citations":[{"cited_title":"Improving the Improved Training of Wasserstein GANs: A Consistency Term and Its Dual Effect","cited_arxiv_id":"1803.01541","evidence_quote":"Provides the CTGAN consistency-regularized Wasserstein GAN loss adopted for the local generators."},{"cited_title":"Gossiping gans : Position paper,","cited_arxiv_id":null,"evidence_quote":"Supplies the gossip-learning model averaging operation (Eqs. 15-16) that merges generator and discriminator weights."},{"cited_title":"DeepMIMO: A generic deep learning datas et for mil- limeter wave and massive MIMO applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepMIMO ray-tracing dataset and channel model used for the two evaluation scenarios."},{"cited_title":"Multi-resolution CSI feedb ack with deep learning in massive MIMO system,","cited_arxiv_id":null,"evidence_quote":"Provides the CRNet multi-resolution architecture used to test compatibility of the Gossip-GAN training."},{"cited_title":"User-centric onli ne gossip training for autoencoder-based CSI feedback,","cited_arxiv_id":null,"evidence_quote":"Supplies the user-centric gossip training idea for autoencoder CSI feedback that motivates the distributed strategy."},{"cited_title":"ChannelGAN: Deep l earning- based channel modeling and generating,","cited_arxiv_id":null,"evidence_quote":"Establishes that GAN-generated channels match the real channel distribution, supporting the use of a generator for synthetic CSI."}],"review_version":1}