{"id":"d5d4bf79-a992-410e-946b-ac0904c683ef","arxiv_id":"2501.16951","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Two federated learning frameworks for multicell ISAC beamforming are proposed: a VFL approach with a central server and an HFL approach with a fully distributed leakage-penalty loss function.","lead":"The paper proposes two federated learning frameworks for coordinating beamforming across base stations in a network that simultaneously communicates with users and senses targets. Simulated results suggest the distributed method approaches the performance of centralized designs while avoiding the exchange of channel state information online.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HFL's global ICI control claim rests on an unproven surrogate-to-global equivalence with undisclosed leakage weights α,β; the near-WMMSE results could be a tuning artifact rather than robust distributed coordination.","rationale":"In good faith, the paper proposes a plausible decentralized coordination scheme: each BS penalizes the interference it leaks to other cells, so jointly minimizing those penalties can reduce global ICI. The simulations are non-trivial and the idea is coherent. I do not see an internal contradiction that would justify REJECT. However, the headline claim is only as strong as the relationship between the per-cell surrogate losses and the true global objective. Equations (21)-(22) are not equivalent to (9), because true SINR includes cross-cell interference while the local losses assume it away and instead add leakage penalties controlled by unreported α and β. The paper states in Section IV-B that minimizing local losses minimizes the global loss, but offers no proof or numerical demonstration that the surrogate and global objective move together. The reader's CONDITIONAL verdict already identifies this as a central weakness; my concern sharpens it by pointing to the specific structural mismatch and the need for an α,β sensitivity analysis. If the proposed check passes, the claim becomes credible; if not, the reported near-WMMSE gains cannot be distinguished from a tuned heuristic. No change to the reader's verdict is needed, but the conditions should include reporting α,β and adding robustness evidence.","tokens_in":17386,"tokens_out":12368,"duration_ms":126039,"concrete_test":"Re-run the HFL experiment from Fig. 5(a) at SNR=20 dB over α,β ∈ {0, 0.01, 0.1, 1, 10, 100}, keeping ρ and all other settings fixed, and record Rc, Rs, and the true global objective (9). Include α=β=0, which reduces to per-cell DL. Also track the surrogate (25) and the true global objective on a fixed validation set during training. If HFL stays near WMMSE across the grid and the true objective tracks the surrogate, the concern fails; if near-WMMSE performance occurs only at one isolated (α,β) point or the true objective drops while the surrogate falls, the claim of robust global ICI control is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the claim in Section IV-B that minimizing the per-cell losses (21)-(22) minimizes the global objective (9). These surrogates do not match (9): true SINR in (3) contains ICI from other BSs, while (21)-(22) replace it with an interference-free local SINR plus a leakage penalty weighted by α and β. Minimizing the sum of these local losses is therefore not equivalent to maximizing ρRc+(1−ρ)Rs; it is a heuristic that requires α and β to be strong enough to suppress leakage but not so strong that the beamformers over-null and sacrifice local rates. The paper provides no bound linking (25) to (9), and Section V never discloses α and β or studies their sensitivity. Since the central claim is 'global interference control using only local channel information' with performance 'comparable to centralized methods,' and all evidence comes from simulations of this surrogate, the missing leakage weights and absent robustness analysis are load-bearing. The observed performance could be an artifact of a narrow, carefully chosen operating point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two federated-learning-based beamforming frameworks for multi-cell ISAC systems. The first, VFL-based, has base stations train local DNNs while a central server computes the global communication and sensing loss and feeds it back for local model updates. The second, HFL-based, is fully decentralized: each BS trains a shared model locally using a newly proposed interference-leakage-based loss function (21)-(22), with FedAvg aggregation. The authors claim both methods can manage inter-cell interference using only local channel information and achieve performance comparable to centralized WMMSE while significantly reducing communication overhead and computational complexity. Numerical simulations compare the proposed schemes against WMMSE, per-cell deep learning, MRT, IMT, and CBF benchmarks under various SNRs, antenna counts, and S&C tradeoff weights, and also study model pruning.","tokens_in":17579,"tokens_out":8015,"duration_ms":70242,"significance":"If validated, the paper would provide a practical low-overhead alternative to centralized coordinated beamforming in multi-cell ISAC, and the HFL interference-leakage loss is an interesting and potentially useful construction. The paper also gives a detailed complexity and communication-overhead analysis and includes pruning experiments, which are valuable for deployment considerations. However, the current manuscript has several load-bearing technical gaps: the VFL training procedure as described is not implementable, the sensing SINR in Eq. (7) appears inconsistent with Eq. (6), and the HFL surrogate-to-global equivalence relies on undisclosed hyperparameters. These issues must be resolved before the central claims can be considered established.","major_comments":[{"comment":"The sensing SINR in Eq. (7) appears to have a factor-N_R inconsistency relative to Eq. (6). After MRC combining with v_m = b(θ_m), the signal component in Eq. (6) is N_R α_m a^H(θ_m) W_m s_m, so the signal power in the SINR expression should scale as N_R^2 ∑_k |g^H_{m,m} w_{m,k}|^2 (with g_{m,m} = α_m a(θ_m)), whereas Eq. (7) uses only a single factor N_R in the numerator and keeps σ_s^2 in the denominator. This changes the numerical radar information rates by a factor of N_R (about 7.8 dB for N_R=6) and may therefore alter the reported sensing performance and the conclusions drawn from Figs. 5(b) and 8.","section":"Section II-B, Eq. (7)"},{"comment":"The VFL backward-propagation step is not well-defined. The paper states that each BS receives the global loss scalar and then performs individual backpropagation using the chain rule (Eq. (18)). However, the gradient ∂ℓ/∂ω_m of the global loss with respect to local parameters depends on ∂ℓ/∂W_m, which requires knowledge of all cross-cell channels (e.g., h_{n,m,k} for n≠m) and the beamformers of other BSs. A scalar global loss value does not provide this derivative information. To make the VFL procedure implementable, the server would need to transmit the gradient of the global loss with respect to each local output matrix W_m (or an equivalent error signal) to each BS, or the paper must explain how the local BS can compute the gradient without this information.","section":"Section III-C, Step 4 and Eq. (18)"},{"comment":"The central claim for the HFL framework rests on the assertion in Section IV-B that minimizing the local losses (21)-(22) leads to the minimization of the global loss (25), and hence solves the original problem (9). The local losses replace the actual SINR expressions, which contain inter-cell interference, with interference-free local SINRs plus leakage penalties weighted by α and β. There is no analytical characterization of the relationship between the surrogate (25) and the original objective (9); the paper only states that simulations verify the effectiveness. Moreover, the values of α and β used in Section V are never disclosed, and no sensitivity analysis is provided. Without these, the reported 'performance comparable to centralized methods' could be a tuning artifact of unreported hyperparameters rather than a robust property of the proposed method.","section":"Section IV-B, Eqs. (21)-(22) and (25)"}],"minor_comments":[{"comment":"In Eq. (21), the CIL penalty term is written as ∑_{n≠m} ∑_{i=1}^K |h^H_{m,n,i} w_{m,k}|^2, which is inconsistent with the definition of CIL in Eq. (19) where the beamforming matrix W_m appears. As written, the index k is free and the sum over k is missing. Please correct the expression, for example to ∑_{n≠m} ∑_{i=1}^K ∑_{k=1}^K |h^H_{m,n,i} w_{m,k}|^2.","section":"Section IV-A, Eq. (21)"},{"comment":"The line 'W∗_m = f([H_m, G_m]; ω∗), ∀m ⊂ M' uses the subset symbol ⊂ instead of the membership symbol ∈; the correct form is ∀m ∈ M.","section":"Algorithm 1 and Algorithm 2"},{"comment":"There is a notation inconsistency between ω and w: Eq. (18) writes ω_m^{(T+1)} = w_m^{(T)} − η∇ℓ, where the right-hand side uses w instead of ω. This should be ω_m^{(T)} for clarity.","section":"Section III-C, Eq. (18)"},{"comment":"In the provided text, Table II is not formatted correctly: the HFL row contains running text instead of a complexity expression, and the table appears incomplete. Please provide a clean table with the computational complexities for all three methods.","section":"Table II"},{"comment":"The benchmarks are not fully specified: for WMMSE, the authors cite [16] but do not describe how the communication-only WMMSE is extended to the ISAC objective (9) with the sensing rate; for the per-cell DL benchmark in Fig. 5(b), the network architecture and training details are not given. Adding these details would improve reproducibility.","section":"Section V-A and V-B"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and relevant problem, and the HFL leakage-based loss is an interesting idea. However, the VFL training description appears conceptually incomplete, the sensing SINR contains a likely factor-N_R error, and the HFL method's key hyperparameters are undisclosed. These are fixable in a revision, so I recommend major revision rather than rejection. I would also encourage the editor to ask the authors to provide a clear derivation or formal justification of the relation between the HFL surrogate and the original problem, or at least an extensive sensitivity study over α and β."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a reasonable application of VFL and HFL to multicell ISAC beamforming, with a useful HFL loss that makes training fully local. The paper is readable and the simulation study is broad. But there are three fixable issues that currently block a clean endorsement: an apparent scaling error in the sensing SINR, an underspecified VFL backward step, and an HFL loss whose key weights (alpha and beta) are never reported. The stress-test concern about HFL is legitimate: the 'global interference control' claim rests on an unproven heuristic with undisclosed tuning knobs.\n\nWhat is actually new: the combination of leakage-based losses with HFL so that each BS trains only on local channels, then FedAvg merges the models; prior work used FL for cell-free precoding or leakage minimization in a centralized manner. The gains over per-cell DL, MRT, and CBF in the simulations are plausible, and the pruning analysis is a nice extra.\n\nWhere it wobbles. First, Eq. (7) has an NR factor in the numerator where Eq. (6) implies NR^2. Given NR=6, the reported radar information rates are off by a non-negligible amount, and the HFL local sensing loss in Eq. (22) omits the same factor, so VFL and HFL sensing objectives are inconsistent. Second, in the VFL training, local BSs need the gradient of the global loss with respect to their own output; the paper merely says the 'global loss is fed back.' That is too vague to reproduce. Third, the HFL loss is explicitly a heuristic -- the paper says it can be verified through simulations -- and alpha and beta are never disclosed, nor is their sensitivity studied. This is the load-bearing soft spot. Without those values, the claim that HFL gives global interference control from local data cannot be independently checked, and the near-WMMSE results could indeed be a carefully tuned operating point. The other sections are standard and the comparison setup is fair.\n\nWho it is for: the ISAC physical-layer learning community. It deserves a serious referee, because the idea is sound and the flaws are correctable. I would not cite it in its current form, but after a revision that fixes the scaling, details the VFL gradient flow, and reports alpha and beta with ablations, it becomes a useful paper.","headline":"A useful FL-for-ISAC idea undermined by an undisclosed HFL weighting and a sensing SINR scaling error; fixable.","tokens_in":18134,"tokens_out":3712,"would_cite":false,"duration_ms":33626,"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":"Federated learning coordinates multi-cell ISAC beamforming without sharing channel data.","keywords":["federated learning","integrated sensing and communications (ISAC)","coordinated beamforming","inter-cell interference","interference leakage","multi-cell networks","deep learning beamforming","vertical federated learning"],"falsifier":"Train the HFL framework in an interference-limited cell-edge scenario with the leakage weights $\\alpha$ and $\\beta$ swept from zero to large values, and compare the resulting global communication-plus-sensing rate against centralized WMMSE; if no setting of $\\alpha$ and $\\beta$ reproduces centralized performance, the claim that local loss minimization controls global interference is falsified.","tokens_in":17184,"feed_emoji":"📡","tokens_out":4202,"duration_ms":36957,"temperature":0.7,"pith_summary":"This paper works on multi-cell integrated sensing and communications (ISAC), where base stations share spectrum for both communication and radar sensing. The problem is that beamforming from local channel knowledge alone creates inter-cell interference that hurts both users and neighboring-cell sensing receivers, while centralized beamforming needs global channel information and costly backhaul exchange. The authors propose two federated-learning designs that train deep networks for beamforming across base stations: a vertical-federated-learning (VFL) scheme in which a central server computes a global loss from locally designed beamformers, and a horizontal-federated-learning (HFL) scheme in which each base station trains solely on its own channel data using a loss that penalizes interference leakage. The central claim is that these distributed designs can reach performance comparable to centralized methods while keeping channel data local, cutting communication overhead and latency. If true, coordinated multi-cell ISAC beamforming becomes feasible in dense and latency-sensitive deployments.","feed_headline":"Federated beamforming cuts inter-cell interference with no global CSI","feed_subtitle":"Two distributed training schemes match centralized performance while keeping channel data local, cutting overhead and latency.","key_machinery":"The load-bearing objects are the local loss functions in equations (21) and (22): the communication loss $L_c(W_m)$ adds a weighted penalty $\\alpha$ on the total power leaked from base station $m$ toward all unintended users in other cells, and the sensing loss $L_s(W_m)$ adds a weighted penalty $\\beta$ on the power leaked toward neighboring base stations' sensing receivers. Because each station can compute these leakage terms from its own channel estimates, the global problem splits into per-cell problems that FedAvg (averaging model parameters across stations) combines into a shared network. This leakage-based formulation is what lets the authors claim a fully decentralized training phase with no channel-data exchange.","core_discovery":"The paper's central claim is that global interference control in multi-cell ISAC does not require global channel state information. The VFL framework trains each base station's local network branch while a central server evaluates the true global communication and sensing losses and feeds them back, so the stations learn to shape beams that suppress interference to other cells without exchanging channel data online. The HFL framework goes further: each station minimizes a local loss composed of its own achievable rates plus penalty terms for communication interference leakage and sensing interference leakage, defined via local channel estimates to neighboring receivers. With FedAvg-style aggregation of the model weights, the paper argues that minimizing these per-cell losses drives the global network objective to a near-optimal point, and numerical results show the HFL design closing most of the gap to centralized WMMSE while outperforming per-cell deep-learning and closed-form beamformers, especially in interference-limited regimes.","pith_inferences":["Inference: extending the leakage-loss idea to time-varying channels, one could periodically recompute the penalty terms from fresh local CSI and retrain only the affected base station, a possibility the paper does not explore.","Inference: because the leakage penalties only suppress power leaked along known interfering directions, the HFL performance should degrade when the interfering channels are estimated with error; a robustness study would test this boundary.","Inference: the same local-loss construction could transfer to other coordinated radio problems, such as pilot assignment or power control in dense cell-free networks, wherever inter-node interference can be expressed as leaked power."],"forward_implications":["In interference-limited deployments, the HFL design should deliver most of the interference-nulling gain of centralized coordination while requiring only local channel estimates.","Communication overhead during training drops from uploading channel samples and beamformers (VFL) to uploading only model parameters (HFL), enabling fine-tuning or retraining in dense networks.","The same local leakage-penalty idea can be applied when the served users or targets move, since each station can recompute its leakage terms from its own updated channels.","Model pruning can remove 50–60% of network parameters while keeping performance near the unpruned model, making the beamforming network light enough for practical base stations."],"supporting_citations":[{"why":"It supplies the FedAvg algorithm used for global model aggregation in the HFL framework.","marker":"[26]"},{"why":"It provides the WMMSE centralized benchmark whose performance the proposed methods are compared against.","marker":"[16]"},{"why":"It supplies the per-cell unsupervised deep-learning beamforming baseline and the loss-function style the paper adapts.","marker":"[21]"},{"why":"It supplies the stochastic-geometry model showing that inter-cell interference dominates network-level sensing performance.","marker":"[12]"},{"why":"It establishes the HFL/VFL precoding approach for cell-free systems that this paper extends to multi-cell ISAC.","marker":"[33]"},{"why":"It provides the interference-leakage workaround used to define the communication and sensing leakage penalties.","marker":"[41]"},{"why":"It supplies the FedAvg convergence result used to characterize the number of communication rounds for the HFL framework.","marker":"[42]"},{"why":"It provides the WMMSE complexity expression that supports the paper's complexity comparison.","marker":"[40]"}],"fun_headline_variants":["Two federated schemes tame inter-cell interference without global CSI","Distributed beamforming rivals centralized ISAC without sharing CSI","FL-based beamforming matches centralized gains with zero CSI exchange","Two FL approaches cut overhead, match centralized ISAC beamforming"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that independently minimizing each base station's local loss, with its unreported leakage weights, pushes the global network objective to a near-optimal point, and that averaging the locally trained model weights converges to that solution.","fun_headline_variants_meta":{"raw":{"variants":["Two federated schemes tame inter-cell interference without global CSI","Distributed beamforming rivals centralized ISAC without sharing CSI","FL-based beamforming matches centralized gains with zero CSI exchange","Two FL approaches cut overhead, match centralized ISAC beamforming"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000604,"raw_usage":{"total_tokens":2838,"prompt_tokens":982,"completion_tokens":1856,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":1788}},"tokens_in":598,"tokens_out":1856,"duration_ms":10934,"temperature":1.0,"reasoning_tokens":1788,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T05:30:46.967870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the HFL framework in an interference-limited cell-edge scenario with the leakage weights $\\alpha$ and $\\beta$ swept from zero to large values, and compare the resulting global communication-plus-sensing rate against centralized WMMSE; if no setting of $\\alpha$ and $\\beta$ reproduces centralized performance, the claim that local loss minimization controls global interference is falsified.","supporting_citations":[{"cited_title":"An iteratively weighted mmse approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,","cited_arxiv_id":null,"evidence_quote":"It provides the WMMSE centralized benchmark whose performance the proposed methods are compared against."},{"cited_title":"Unsu- pervised learning-based fast beamforming design for downlink MIMO,","cited_arxiv_id":null,"evidence_quote":"It supplies the per-cell unsupervised deep-learning beamforming baseline and the loss-function style the paper adapts."},{"cited_title":"Network-level inte- grated sensing and communication: Interference management and BS coordination using stochastic geometry,","cited_arxiv_id":null,"evidence_quote":"It supplies the stochastic-geometry model showing that inter-cell interference dominates network-level sensing performance."},{"cited_title":"Federated learning for precoding design in cell-free massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"It establishes the HFL/VFL precoding approach for cell-free systems that this paper extends to multi-cell ISAC."},{"cited_title":"Multi-cell coordinated joint sensing and communications,","cited_arxiv_id":null,"evidence_quote":"It provides the interference-leakage workaround used to define the communication and sensing leakage penalties."},{"cited_title":"Iterative algorithm induced deep-unfolding neural networks: Precoding design for mul- tiuser MIMO systems,","cited_arxiv_id":null,"evidence_quote":"It provides the WMMSE complexity expression that supports the paper's complexity comparison."}],"review_version":1}