{"id":"29d8d364-45b9-4135-a894-1497e1254351","arxiv_id":"2608.00855","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fairness-consensus bilevel optimizer that adjusts transmission thresholds and powers improves packet delivery and federated learning accuracy for UAV-enabled IoT networks in simulations.","lead":"A new transmission controller for drone-served federated learning decides when devices send packets and how much power they use, and in simulations it raises the share of updates that arrive intact from under half to ninety percent. The paper shows that accounting for packet-level losses, rather than assuming perfect delivery, changes how federated learning should be scheduled over unlicensed wireless bands.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FL gains rest on an unvalidated analytical PDR model and independent-Bernoulli packet loss; the same model is used for optimization and evaluation, so a biased R_n would invalidate the reported improvements.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the packetized FL aggregation relies on independent Bernoulli draws with success probability R_n from Eq. (10), which depends on log-normal interference parameters not provided and on independence assumptions not justified. This is the most critical point for the central claim because the entire empirical demonstration is a closed loop: FCB optimizes the analytical PDR model, and the same analytical model generates the packet-loss masks used to evaluate FL training. If the model is biased or losses are correlated, the reported PDR gains and CNN accuracy improvements are artifacts of the simulation, not evidence about real UAV-enabled FL. The circularity in Eq. (17f) is also a real formal flaw, but it affects the claimed contribution as an optimization problem rather than the empirical comparison itself; the implemented algorithm can still be evaluated heuristically. I agree with the reader that this weakness warrants a CONDITIONAL verdict, not outright rejection, because the core idea (co-designing thresholds and powers for PDR-aware aggregation) is salvageable and the missing pieces are identifiable: provide μ,σ, test the independence assumption, and validate against a packet-level simulator or testbed. My concrete test directly targets the independence and model-bias assumptions, and if it passes, the central claim would be substantially strengthened.","tokens_in":10434,"tokens_out":5781,"duration_ms":53193,"concrete_test":"Re-run the FL experiments with a packet-level simulator (e.g., ns-3 or a custom Gilbert-Elliott channel) using the β and P from Table I, replacing independent Bernoulli draws with correlated loss processes whose marginal probabilities match the reported R_n values. Also recompute R_n using the μ,σ from [5] (or estimate them by Monte Carlo interference simulation) and retrain the CNN under both IID and non-IID splits. If FCB's advantage over the conservative and aggressive baselines shrinks or reverses, the Bernoulli/PDR assumption is load-bearing; if the relative ordering persists, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (10) approximates PDR as R_n ≈ [1 - P_d^n - P_o^n - P_e^n]_0^1, where P_e^n depends on log-normal interference parameters μ(β_-n,P_-n) and σ(β_-n,P_-n) that are 'as derived in [5]' and never specified in this paper. Consequently, R_n cannot be independently reproduced or checked. The FL simulation then treats each packet indicator b_{n,c}^{(t)} as an independent Bernoulli draw with success probability R_n (Sec. II-B), and Eq. (16) aggregates masked updates. This independence assumption is not justified: block fading and queueing induce temporal and cross-packet correlation; an interference burst can drop an entire C-packet update in one round, which the i.i.d. Bernoulli model cannot represent. Moreover, the same analytic R_n is maximized by CTC/FPC and then used as the ground truth in the FL evaluation, making the comparison in-sample: FCB appears superior because it maximizes the very function used to simulate the channel, not because of demonstrated real-world packet delivery. Since μ, σ, P_d, and P_o are all imported from prior work without sensitivity analysis, error bars, or testbed validation, the central claim that FCB improves FL aggregation and training by enhancing packet-level update delivery is unsupported if the PDR model is biased or if packet losses are correlated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a packet-level transmission framework for UAV-enabled federated learning (FL) in unlicensed ISM bands, modeling packet loss due to buffer overflow, delay violations, and interference-dependent transmission errors. A packet delivery ratio (PDR) model from prior work is used to represent partial update reception as independent Bernoulli-masked packet losses in FL aggregation. The authors formulate a fairness-consensus bilevel (FCB) optimization problem that jointly tunes transmission thresholds (inner level) and transmission powers (outer level), and propose an alternating optimizer consisting of a consensus-based threshold controller (CTC) and a fairness-based power controller (FPC). Numerical results on CIFAR-10 report that FCB achieves higher average PDR and Jain's fairness index than two baselines, and that these gains translate into faster convergence and higher FL accuracy under IID and non-IID data.","tokens_in":10741,"tokens_out":4157,"duration_ms":39943,"significance":"If the results are robust, the paper addresses a real gap in wireless FL by explicitly modeling packet-level partial update reception and coupling it with transmission control, rather than assuming error-free scheduled updates. The proposed framework is relevant to UAV-assisted edge intelligence in shared spectrum, and the idea of jointly optimizing threshold and power for PDR fairness under partial observability is useful. The paper also gives credit for including a packetized Bernoulli-masked aggregation model and for demonstrating the performance gap between ideal and lossy uplinks. However, the significance is currently limited by validation gaps: the PDR model is imported from prior work without enough detail to reproduce it, the optimization problem is not well-posed as written, no convergence guarantees are provided, and the FL simulations lack statistical error bars.","major_comments":[{"comment":"The PDR expression R_n(β,P) is the cornerstone of the paper, but it is not self-contained: the log-normal interference parameters μ(β_-n,P_-n) and σ(β_-n,P_-n) in Eq. (8) are said to be 'as derived in [5]' and are never specified. A reader cannot compute or independently check R_n, P_e^n, or any downstream result. More importantly, the same R_n is both the objective that CTC/FPC maximize (Eq. (17)) and the ground-truth success probability used to generate Bernoulli-masked FL updates (Eq. (15)). This makes the evaluation in-sample: any bias or approximation error in the imported PDR model is automatically reflected as an apparent FL improvement. The authors should provide explicit expressions or a self-contained derivation for μ, σ, P_d, and P_o, and validate the PDR model against packet-level simulation or a testbed, including a sensitivity analysis over the unspecified parameters.","section":"Section II-A, Eq. (10), and Section II-B"},{"comment":"The inner-level optimization problem is not well-posed as written. Equation (17d) defines β⋆ as the arg max of the average PDR, but the constraint (17f) then requires β_n - β⋆[n] = 0, so the feasible set depends on the very vector being optimized. This makes the formulation self-referential and prevents any direct interpretation of the inner problem. The intended consensus condition should be expressed with an explicit consensus variable, such as β_n = β_c for a scalar consensus value β_c or an equivalent constraint formulated before the arg max is taken. The current formulation must be repaired before the CTC algorithm can be said to solve the stated problem.","section":"Section III-A, Eqs. (17d)-(17f)"},{"comment":"No convergence analysis is given for the CTC, FPC, or the alternating FCB optimizer. CTC relies on random participation with a decaying drop probability ψ_i and a heuristic local coordinate search, while FPC uses greedy coordinate-wise power updates and a Jain's index correction (JIC) subroutine that is described only informally. The paper states that FCB converges to the reported operating point, but it does not show that CTC reaches a consensus maximizer, that FPC improves the worst PDR monotonically, or that the alternating procedure converges even to a local solution. The authors should provide a formal convergence or monotonicity argument, or clearly state the conditions under which the algorithms are guaranteed to terminate.","section":"Section III-B, Algorithms 2 and 3"},{"comment":"The FL aggregation model assumes that each packet reception indicator b_{n,c}^{(t)} is an independent Bernoulli draw with success probability R_n. This independence assumption is not justified in a block-fading, interference-coupled ISM channel. Correlated losses caused by interference bursts or queueing dynamics can drop entire multi-packet updates in a round, which the i.i.d. Bernoulli model cannot capture. Since Eq. (16) aggregates masked updates exactly according to this model, the reported FL accuracy gains depend on the validity of the independence assumption. The authors should either justify the assumption from the channel model or evaluate robustness under a correlated-loss model, such as a Gilbert-Elliott or block-fading packet-loss process.","section":"Section II-B, Eq. (15)"},{"comment":"The FL training results in Fig. 2 are presented as single curves with no error bars, confidence intervals, or number of independent seeds. Given that the aggregation process is stochastic (Bernoulli masking with random seeds), the claimed differences in convergence speed and final accuracy—such as the 79% improvement in the strong non-IID case—could be within run-to-run variability. The authors should repeat the FL experiments over multiple seeds and report means and standard deviations or confidence bands.","section":"Section IV"}],"minor_comments":[{"comment":"The manuscript frequently types the acronym as 'UA V' instead of 'UAV'; this should be corrected throughout.","section":"Throughout"},{"comment":"The figure captions read 'CIFR-10' in three places; the dataset name is CIFAR-10.","section":"Fig. 2"},{"comment":"The queueing model parameter λ_n = 100 is given without units; specifying packets per time slot or packets per second would improve reproducibility.","section":"Table II"},{"comment":"The notation for the best sub-channel f⋆ is introduced in the text, but the exponent |F| in Eq. (5) implicitly assumes independent fading across sub-channels; stating this assumption explicitly would be helpful.","section":"Section II-A, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The paper's core PDR model is imported from the authors' prior work [5] without the parameter values needed for independent computation, and the same model is used for both optimization and evaluation. This creates a reproducibility risk that the editorial board should weigh when deciding whether the paper meets the journal's standards. If the authors can provide a self-contained derivation or supplementary code, this concern would be substantially alleviated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of arXiv:2608.00855. The paper is worth engaging with, but the main quantitative claims rest on a single analytical PDR model taken from the authors' prior work [5], and that same model does double duty as the optimizer's objective and the simulation's channel. If that model is biased, or if the independent-Bernoulli packet-loss assumption fails, the reported gains shrink accordingly.\n\nWhat's actually new: the FCB optimizer, a bilevel scheme that couples transmission thresholds (consensus) with transmission powers (fairness) for packetized federated learning. The Bernoulli-masked aggregation in Eqs. (15)-(16) is a natural way to turn PDR into partial-update reception, and it addresses a real gap in the wireless-FL literature, which usually assumes scheduled updates arrive whole. The numerical results show faster convergence and better final accuracy under non-IID data, and the authors are transparent that the communication model comes from [5]. Building on prior work is legitimate, not a flaw per se.\n\nThe soft spots are real but not fatal. The inner constraint (17f) is circular as written: β⋆ is defined as the argmax, then each β_n is forced to equal β⋆[n], making the constraint self-referential. The algorithm is an alternating heuristic without a convergence proof—common for this type of problem, but it should be stated. More importantly, the evaluation is in-sample: FCB maximizes R_n, and the FL simulation draws packet success from that same R_n. The FL accuracy metric is distinct, so this is not direct circularity, but any bias in R_n flows straight into the learning results. The log-normal interference parameters μ and σ are \"as derived in [5]\" and never specified, so an independent reader cannot reproduce or sensitivity-check the PDR. I'd also like error bars over network realizations, since only one PPP draw appears to be reported.\n\nThe independence assumption on packet losses is stated explicitly (\"standard assumption of independent packet transmissions\") but not defended. In a real ISM band with interference bursts, entire C-packet updates can drop in a round; a Bernoulli mask with mean R_n will not capture that correlation. Whether that correlation matters for FL convergence is an empirical question, and the paper doesn't address it.\n\nMy verdict: conditional accept for a venue, but only if the authors provide the missing model parameters, release code/data, add sensitivity analysis, and fix the inner-problem statement. The core idea—co-designing thresholds and powers around packet-level FL delivery—is solid and worth refereeing. The reader's take is roughly right, though I'd weight the in-sample concern more heavily than the circularity concern.\n\nRecommendation: send it to serious peer review, with a strong request for code and sensitivity analysis.","headline":"Worth engaging: the FCB co-design of thresholds and powers for packetized UAV-FL is new and relevant, but the numerical claims are only as strong as an unvalidated analytical PDR model that serves as both the optimizer's objective and the simulation's channel.","tokens_in":11236,"tokens_out":2872,"would_cite":true,"duration_ms":27200,"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":"This paper claims that a fairness–consensus bilevel transmission controller, built on a packet-level PDR model, achieves 0.90 average packet delivery and 0.99 fairness in UAV-enabled federated learning and outperforms aggressive and…","keywords":["federated learning","UAV-enabled networks","packet delivery ratio","Bernoulli masking","transmission threshold control","power fairness","partial observability","non-IID data"],"falsifier":"A direct test is to replace the independent-Bernoulli mask with empirical erasure traces from a real UAV-to-IoT link or a correlated-interference simulator; if packet losses at a fixed mean PDR are positively autocorrelated, or if the empirical PDR deviates from Eq. (10)'s log-normal prediction, the reported convergence and accuracy gaps are not reproducible.","tokens_in":10215,"feed_emoji":"📡","tokens_out":6757,"duration_ms":56977,"temperature":0.7,"pith_summary":"Federated learning over a UAV in unlicensed bands is usually analyzed as if scheduled model updates arrive intact; this paper argues that uplink updates are actually packetized and can be partially lost to buffer overflow, delay violations, and transmission errors. It builds a packet-level delivery model whose key output, the packet delivery ratio $R_n(\\beta,P)$, enters federated aggregation as a Bernoulli mask: each packet of an update is kept or dropped independently with that probability. On top of this, the paper formulates a fairness-consensus bilevel optimization that first drives per-learner transmission thresholds to a PDR-efficient consensus and then adjusts transmission powers to raise the worst PDR under a Jain's fairness constraint. Numerical CNN training on CIFAR-10 reports that the resulting FCB controller reaches 0.90 average PDR and 0.99 fairness, converges faster in IID settings, and substantially beats aggressive and conservative policies as data become non-IID. If correct, this gives a concrete way to design transmission control from the learning objective, closing part of the gap between idealized wireless-FL assumptions and practical partial-update reception.","feed_headline":"Transmission tuning lifts UAV federated learning packet delivery to 90%","feed_subtitle":"A threshold-and-power controller beats aggressive and conservative baselines on delivery and CIFAR-10 accuracy.","key_machinery":"The load-bearing object is the packetized PDR model $R_n(\\beta,P)\\approx[1-P^d_n(\\beta_n)-P^o_n(\\beta_n)-P^e_n(\\beta,P)]_0^1$, where the three subtracted terms are delay-violation, buffer-overflow, and transmission-error probabilities derived from Nakagami-m fading, queueing, and log-normal interference. That scalar is used as the success probability of an independent Bernoulli mask $b^{(t)}_{n,c}$ that zeros out lost packets in each learner's update before global aggregation. Two coordinated controllers carry the optimization: the consensus-based threshold controller (CTC) runs local coordinate searches to push thresholds $\\beta$ toward a consensus that maximizes average PDR, and the fairness-based power controller (FPC) reallocates power to raise the minimum PDR while enforcing a weighted Jain's index constraint. Together they alternate through the fairness-consensus bilevel problem stated as (17a)–(17f).","core_discovery":"The paper's central claim is that the transmission policy itself, not just the learning algorithm, can be optimized against a realistic packet-level measure of update delivery, and that doing so restores much of the performance lost to partial reception. Concretely, the FCB optimizer selects moderate transmission thresholds and powers that avoid both the aggressive policy's interference flooding, which yields an average PDR of 0.39 and fairness of 0.80, and the conservative policy's excessive buffering, which yields 0.49 and 0.81. The paper reports that FCB raises average PDR to 0.90 with Jain's index 0.99 under partial observability, keeps per-learner PDR consistently high, and translates these gains into faster convergence on IID CIFAR-10 and roughly 32% and 79% higher final accuracy than the conservative and aggressive baselines under strong non-IID data. The authors summarize this as the FCB optimizer improving FL aggregation and training performance by enhancing packet-level update delivery.","pith_inferences":["If packet erasures are correlated rather than independent, the Bernoulli-mask abstraction would understate the damage of interference bursts; a natural extension is to replace the i.i.d. mask with a Markov or bursty erasure model and re-test FCB.","The same bilevel consensus-fairness design could be applied to other packetized uplink tasks, such as gradient compression with variable packet sizes, where the delivery metric is not raw PDR but useful information delivered per packet.","The paper's implicit claim that PDR is a sufficient statistic for FL aggregation quality suggests a simpler deployment rule: measure per-learner PDR online and use it to adjust transmission parameters, which is directly testable on a software-defined radio testbed."],"forward_implications":["If FCB is correct, serving federated learning from a UAV over unlicensed bands does not require ideal error-free links; a PDR-aware threshold-and-power policy keeps aggregation quality close to the ideal-channel benchmark.","Under IID data, the gain is mostly convergence speed: FCB reaches 55% accuracy in about 45 communication rounds, versus about 65 for the conservative policy and 80 for the aggressive policy.","Under strong non-IID data, the gain becomes final accuracy: FCB reaches around 50%, compared with about 38% for conservative and 28% for aggressive transmission.","Partial observability changes convergence cost rather than the solution: with a learner drop probability of 0.4, the CTC needs 62 inner iterations instead of 10 but converges to the same PDR and fairness point."],"supporting_citations":[{"why":"Supplies the PDR model in Eq. (10) including the log-normal interference parameters, and provides the conservative transmission policy baseline.","marker":"[5]"},{"why":"Gives the threshold-based transmit-or-buffer decision and the aggressive transmission policy baseline.","marker":"[16]"},{"why":"Provides the distance-based LoS probability model used for the ground-to-air channel.","marker":"[14]"},{"why":"Provides the Nakagami-m and Rician fading distributions used for the threshold CDF and transmission error probability.","marker":"[15]"},{"why":"Defines the weighted Jain's index used as the fairness constraint in the FCB problem.","marker":"[17]"},{"why":"Underlies the consensus-based distributed optimization structure used by the CTC algorithm.","marker":"[18]"},{"why":"Supplies the CIFAR-10 dataset used for all FL training evaluations.","marker":"[19]"},{"why":"Provides the Dirichlet distribution used to generate mid and strong non-IID data splits.","marker":"[21]"}],"fun_headline_variants":["Adaptive transmission control boosts UAV FL delivery to 90%","UAV FL: one controller lifts packet delivery to 90% and fairness to 0.99","Threshold and power tuning achieve 90% PDR in UAV federated learning","Boosting UAV FL update delivery: adaptive thresholds and power control hit 90%","How a threshold-power controller lifts UAV FL delivery to 90%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on treating each packet as an independent coin flip whose success probability is the analytically computed PDR; if real packet losses are correlated in bursts or the PDR formula's interference parameters are off, the reported learning gains do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive transmission control boosts UAV FL delivery to 90%","UAV FL: one controller lifts packet delivery to 90% and fairness to 0.99","Threshold and power tuning achieve 90% PDR in UAV federated learning","Boosting UAV FL update delivery: adaptive thresholds and power control hit 90%","How a threshold-power controller lifts UAV FL delivery to 90%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3146,"prompt_tokens":981,"completion_tokens":2165,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2063}},"tokens_in":597,"tokens_out":2165,"duration_ms":13751,"temperature":1.0,"reasoning_tokens":2063,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:16:07.526610+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to replace the independent-Bernoulli mask with empirical erasure traces from a real UAV-to-IoT link or a correlated-interference simulator; if packet losses at a fixed mean PDR are positively autocorrelated, or if the empirical PDR deviates from Eq. (10)'s log-normal prediction, the reported convergence and accuracy gaps are not reproducible.","supporting_citations":[{"cited_title":"Channel-aware distributed transmission control and video streaming in UA V networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the PDR model in Eq. (10) including the log-normal interference parameters, and provides the conservative transmission policy baseline."},{"cited_title":"To transmit or not to trans- mit? distributed queueing games in infrastructureless wireless networks,","cited_arxiv_id":null,"evidence_quote":"Gives the threshold-based transmit-or-buffer decision and the aggressive transmission policy baseline."},{"cited_title":"Impact of an interfering node on unmanned aerial vehicle communications,","cited_arxiv_id":null,"evidence_quote":"Provides the distance-based LoS probability model used for the ground-to-air channel."},{"cited_title":"Performance analysis of RIS-assisted UA V communication in NOMA networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Nakagami-m and Rician fading distributions used for the threshold CDF and transmission error probability."},{"cited_title":"Optimal tradeoff between sum-rate efficiency and jain’s fairness index in resource allocation,","cited_arxiv_id":null,"evidence_quote":"Defines the weighted Jain's index used as the fairness constraint in the FCB problem."},{"cited_title":"Distributed optimization and statistical learning via the alternating direction method of multipliers,","cited_arxiv_id":null,"evidence_quote":"Underlies the consensus-based distributed optimization structure used by the CTC algorithm."},{"cited_title":"Ensemble distillation for robust model fusion in federated learning,","cited_arxiv_id":null,"evidence_quote":"Provides the Dirichlet distribution used to generate mid and strong non-IID data splits."}],"review_version":2}