{"id":"44507022-0d66-4b41-a97d-f8fac77233b7","arxiv_id":"2506.02972","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"An online aerial federated learning algorithm that combines model pruning, probabilistic gradient quantization, and trajectory planning reaches similar simulated accuracy to costly baselines with lower resource use.","lead":"This paper proposes an online federated learning method for aerial vehicles that prunes models and quantizes gradient updates to save computation and communication, while planning flight trajectories to collect fresh data. The CIFAR10 simulation shows test accuracy close to the unmodified baselines, but the convergence proof contains an invalid step and the simulation lacks error bars.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof applies smoothness across two different loss functions in Eq. (23) of the supplement; without a loss-drift bound the convergence guarantee is unsupported.","rationale":"The paper's central claim has three components: the 2CEOAFL algorithm, the theoretical convergence bound (Theorem 1), and the empirical demonstration of comparable accuracy. The proof gap in Eq. (23) directly undermines the theoretical component, which is stated as a key contribution. The issue is not a minor typo: smoothness cannot be applied across two different data distributions, and no loss-drift assumption is supplied anywhere in the main text or supplementary. Even if the first inequality were repaired by adding a loss-drift term, the subsequent 'telescoping' argument would not hold because the global loss function changes each round as D^t changes. Without Theorem 1, the paper's remaining evidence is a single simulation run with no error bars or statistical significance testing, which is too thin to support the claim of 'comparable performance' as a rigorous result. The reader's rejection is therefore appropriate, and our analysis does not change that verdict. We agree with the reader's identification of the load-bearing weakness; the concrete test we propose would settle the matter by forcing an explicit treatment of the loss-drift term. If the authors can supply a valid bound or reformulate the theorem for the online setting, the paper could potentially be reconsidered, but as written the central theoretical claim is unsupported. The empirical results may still be useful as an engineering demonstration, but they do not compensate for the unproven theorem.","tokens_in":25440,"tokens_out":6198,"duration_ms":62950,"concrete_test":"Re-derive Eq. (23) with the loss-function index kept explicit: start from f(w^{t+1}|D^{t+1}) ≤ f(w^t|D^{t+1}) + ⟨∇f(w^t|D^{t+1}), w^{t+1}−w^t⟩ + (β/2)‖w^{t+1}−w^t‖², and attempt to bound the missing term f(w^t|D^{t+1}) − f(w^t|D^t) using Definition 2 and Assumption 1. If no finite assumption-derived bound exists, or if the resulting bound contains the loss drift explicitly and the subsequent telescoping sum is invalid, then Theorem 1 is unproven. This is a purely analytical check requiring no new experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The supplementary proof of Theorem 1 begins with Eq. (23): f(w^{t+1}|D^{t+1}) ≤ f(w^t|D^t) + ⟨∇f(w^t|D^t), w^{t+1} − w^t⟩ + (β/2)‖w^{t+1} − w^t‖², attributed to β-Lipschitz smoothness. Smoothness, however, is a property of a single loss function. For f(·|D^{t+1}) the valid inequality is f(w^{t+1}|D^{t+1}) ≤ f(w^t|D^{t+1}) + ⟨∇f(w^t|D^{t+1}), w^{t+1} − w^t⟩ + (β/2)‖w^{t+1} − w^t‖², with the gradient of the D^{t+1}-loss, not the D^t-loss. The written inequality would require an additional bound on the loss drift f(w^t|D^{t+1}) − f(w^t|D^t), or at least a bound on the gradient difference ⟨∇f(w^t|D^{t+1}) − ∇f(w^t|D^t), w^{t+1} − w^t⟩. Definition 2 bounds squared gradient norms by Φ^t_u, but no assumption bounds loss values themselves or their drift across rounds. Moreover, the first term in (19)/(21), E[f(w^t|D^t)] − E[f(w^{t+1}|D^{t+1})], is treated as a telescoping difference; because the loss function changes with D^t, this term does not telescope over t unless the data distribution is stationary. Thus the central convergence bound is not derived from the stated assumptions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 2CEOAFL, an online aerial federated learning algorithm that combines trajectory-aware data sensing with two resource-saving mechanisms: pruning of the received dense model before local training, and probabilistic stochastic quantization of uploaded gradient differences. The authors model the time-varying spatial data distribution as a Gaussian mixture with time-dependent class weights, solve a per-ACV trajectory optimization problem, and then state Theorem 1, which upper-bounds the expected squared norm of the masked global gradient by a sum of terms attributed to stochastic gradient noise, quantization variance, local data distribution shift, gradient dissimilarity, and model pruning. The empirical section reports CIFAR10/ResNet18 experiments showing that 2CEOAFL achieves test accuracy comparable to the unpruned/unquantized AFL baseline while reducing communication overhead. The proof of the theorem is relegated to an online supplementary file.","tokens_in":25828,"tokens_out":4617,"duration_ms":55020,"significance":"If the theoretical bound were valid, the paper would give a useful decomposition of error sources for online FL with model pruning and gradient quantization, and the trajectory/data-distribution model is an interesting attempt to couple ACV motion planning with continual learning. The empirical comparison against AFL, AFL-Prune, AFL-Quant, and centralized SGD is informative and supports the qualitative claim of communication savings at modest accuracy cost. On the other hand, the paper does not ship code or machine-checked proofs, the theoretical derivation is the central advertised contribution, and the experimental evaluation is a single run without multiple seeds or confidence intervals.","major_comments":[{"comment":"The first inequality of the proof, f(w^{t+1}|D^{t+1}) ≤ f(w^t|D^t) + ⟨∇f(w^t|D^t), w^{t+1} − w^t⟩ + (β/2)‖w^{t+1} − w^t‖², is not justified by Assumption 1. Lipschitz smoothness is a property of a single loss function; for the loss at round t+1 the valid expansion would use f(w^t|D^{t+1}) and ∇f(w^t|D^{t+1}). The written inequality requires an additional bound on the drift of the loss itself, such as f(w|D^{t+1}) − f(w|D^t), as well as a bound on the gradient difference ⟨∇f(w^t|D^{t+1}) − ∇f(w^t|D^t), w^{t+1} − w^t⟩. No such assumption is stated. Consequently the derivation of Theorem 1 is not established from the stated assumptions.","section":"Supplementary Materials, Eq. (23)"},{"comment":"The term E[f(w^t|D^t)] − E[f(w^{t+1}|D^{t+1})] is treated as though it telescopes when the bound is averaged over t. Since the dataset D^t changes every round, this is not a telescoping difference; an extra term of the form E[f(w^{t+1}|D^t) − f(w^{t+1}|D^{t+1})] appears. Without a bound on this loss drift, the time-averaged bound in Eq. (20) is not a convergence bound and the stated θ-suboptimal convergence claim is unsupported. The paper's own Section III-C states that the impacts of newly arrived samples are 'not explicitly modeled,' which is consistent with this gap. The proof must either introduce an explicit loss-drift assumption or reformulate the result as a regret bound with the drift appearing as a non-vanishing error term.","section":"Theorem 1 and Eq. (19)-(20)"},{"comment":"Assumption 5 imposes δ^t_u ≥ ‖w^t_u − w̄^{t,0}_u‖² / ‖w^t_u‖², which ties the pruning ratio to the relative squared difference between the dense and pruned models. This is not a property of the lottery-ticket pruning procedure described in Section III-A, and in the experiments δ^t_u is simply chosen uniformly at random from [0.05, 0.7] with no verification that (17) holds. Since the final pruning-error term in Theorem 1 is derived directly through this assumption, the bound applies only to systems satisfying a condition that the proposed algorithm does not guarantee. The authors should either prove that their pruning rule satisfies (17), measure the condition in the simulations, or replace the assumption with a verifiable bound.","section":"Assumption 5, Eq. (17)"},{"comment":"Even before the proof details, the theorem's bound is not a statement of convergence: the leading term 2(E[f(w^t|D^t)] − E[f(w^{t+1}|D^{t+1})])/(η_t κ) may be negative when the data distribution shifts the loss upward, and the pruning term does not vanish for any positive δ. The authors' Remark 1 acknowledges that the global gradient 'may only converge to a neighborhood of a stationary point.' The paper should state precisely what is being proven: a per-round inequality is not the same as a convergence guarantee, and the notation θ (defined as an average over T) is used in Eq. (20) as if it were a bound that tends to zero, which is not shown.","section":"Section III-B, Theorem 1 statement"}],"minor_comments":[{"comment":"The caption contains a typo: 'CIF AR10' should be 'CIFAR10'.","section":"Figure 3 caption"},{"comment":"The notation '∀t1 := t ∋ [(t+1) mod C = 0]' is confusing; please rewrite the constraint in a clearer form indexing the blocks of C rounds.","section":"Section II-C, constraint C5"},{"comment":"The experimental section reports a single run; adding multiple seeds with mean and standard deviation would substantially strengthen the claim of 'comparable performance.'","section":"Section IV-A"},{"comment":"The definition D^t := ⋃_u D^t_u is correct for the union of local datasets, but the notation f(w^t|D^t) is later used both for the global loss and for the loss evaluated on the time-varying global dataset; please clarify that the global function is the weighted sum defined in Eq. (14).","section":"Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The flaw in Supplementary Eq. (23) is the central load-bearing issue: the proof applies smoothness across two different loss functions and thereby invalidates the main convergence theorem. This is fixable only by adding a loss-drift assumption and reworking the telescoping argument, which is a substantial revision. If the authors cannot supply such an assumption, they should remove the 'convergence bound' claim and present the result as an error decomposition conditional on an unverified pruning condition. The paper is also quite similar in structure to references [9] and [24] from the same group; the editor may wish to check the novelty statement against those works."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one genuinely useful thing: it puts together model pruning, QSGD-style quantization, and trajectory planning for online federated learning with aerial vehicles, and shows in simulation that the combined scheme lands within a couple points of the unpruned, unquantized baseline. The engineering parts are clear, the resource accounting is straightforward, and the limitations section is honest.\n\nThat said, the main theoretical contribution—Theorem 1—has a load-bearing proof gap. In the supplementary, Eq. (23) applies β-smoothness as f(w^{t+1}|D^{t+1}) ≤ f(w^t|D^t) + ⟨∇f(w^t|D^t), w^{t+1}−w^t⟩ + ... . Smoothness is a property of one loss function. The valid inequality would use f(w^t|D^{t+1}) and its gradient. To bridge that, you need a bound on the loss drift between D^t and D^{t+1}; the paper only bounds gradient norms via Φ^t_u, which doesn't imply a loss drift bound. The same issue breaks the telescoping of the first term in (19), since the loss function changes each round. The theorem is not established as written.\n\nThe empirical section is thin: a single run, no error bars, no multiple seeds, and the final accuracy differences between 2CEOAFL and AFL are small but the paper claims \"comparable\" without any statistical support. The pruning ratio is chosen randomly, and the paper's own limitations paragraph concedes that pruning ratios and new-sample impacts are not optimized. These are minor if the theory holds, but they compound the proof problem.\n\nWho should read this: people building resource-efficient federated learning for UAVs will find the algorithm and the experimental setup a useful reference point. The theoretical part needs substantial repair before it can be trusted.\n\nI would not cite this as it stands, but I'd send it to peer review with a strong request for major revision: fix the theorem by adding a proper loss-drift assumption or restate the result as an empirical observation, and rerun the experiments with multiple seeds and confidence intervals. If the authors can do that, this could be a solid systems paper.","headline":"A systems recipe that works in simulation, but Theorem 1's proof is invalid as written and the empirical support is too thin to compensate.","tokens_in":26383,"tokens_out":3552,"would_cite":false,"duration_ms":42564,"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":"Resource-limited drones can prune models and quantize gradients in online federated learning with little accuracy loss.","keywords":["aerial federated learning","online federated learning","model pruning","gradient quantization","trajectory optimization","time-varying data distributions","convergence analysis","resource-constrained UAVs"],"falsifier":"Run 2CEOAFL on a synthetic non-stationary task where the data distribution shifts sharply at a known round, and check whether Theorem 1's inequality (19) holds at that round using measured values; if the expected squared masked-gradient norm exceeds the right-hand side while all stated assumptions hold, the loss-drift step in the proof fails.","tokens_in":25156,"feed_emoji":"🚁","tokens_out":9514,"duration_ms":94359,"temperature":0.7,"pith_summary":"This paper tries to show that online federated learning can be made practical for battery- and bandwidth-limited aerial vehicles without sacrificing model quality. The proposed 2CEOAFL algorithm has each aerial client prune the received dense model into a shallow winning ticket, train only that pruned model, and then probabilistically quantize the accumulated gradient before uploading it to the central server. Trajectories are planned so each vehicle collects fresh data from clusters whose time-varying importance it estimates from a Gaussian mixture model. The main theoretical result, Theorem 1, bounds the expected squared norm of the masked global gradient by a sum of error terms from stochastic gradients, quantization, local data distribution shift, gradient dissimilarity, and a pruning term that only disappears when nothing is pruned. Simulations on CIFAR-10 with ResNet-18 show the pruned-and-quantized scheme reaching test accuracy within a few points of the full unpruned, unquantized federated baseline while using less computation and communication.","feed_headline":"Pruned, quantized drone federated learning keeps accuracy","feed_subtitle":"Aerial vehicles with tight batteries and bandwidth train shallow models and send compressed gradients, matching the full-model baseline.","key_machinery":"The load-bearing object is the masked global gradient $\\overline{\\nabla} f(w^t|D^t) := \\sum_{u=0}^{U-1} \\alpha_u \\nabla f_u(w^t|D^t_u) \\odot m^t_u$, whose expected squared norm Theorem 1 bounds. The binary mask $m^t_u$ records which weights survive pruning at vehicle $u$ in round $t$; it is what makes the trained model shallow, and it is also what keeps the pruning error alive in the bound. The proof combines this masked gradient with the probabilistic upload rule $\\Pi^t_u = d^t_u$ with probability $q^t_u$ and $\\Pi^t_u = Q(d^t_u)$ otherwise, where $Q$ is the low-precision stochastic quantizer of Definition 1 and $d^t_u$ is the accumulated pruned-model gradient difference. The machinery is a standard smoothness-and-variance decomposition: $\\beta$-Lipschitz smoothness turns the loss change into gradient inner products, the bias-free assumptions strip the quantization and mini-batch randomness, and the distribution-shift assumption $\\|\\nabla f_u(w|D^{t-1}_u)-\\nabla f_u(w|D^t_u)\\|^2\\le\\Phi^t_u$ converts the online data drift into one of the error terms. The trajectory model is a separate supporting mechanism: a mixed-integer convex program chooses cluster visits according to a time-varying class distribution $\\psi^t_{u,c}$ normalized from $M_u z_u(t)$.","core_discovery":"The paper's central claim is that a fleet of independently owned aerial connected vehicles can run online federated learning under tight onboard constraints by combining three ideas: selfish trajectory optimization that follows each vehicle's time-varying data distribution, model pruning via the lottery-ticket hypothesis so each vehicle trains only a small subnetwork, and probabilistic low-precision quantization of the uploaded gradient differences. Under the stated assumptions, Theorem 1 upper-bounds the expected squared norm of the masked global gradient $\\overline{\\nabla} f(w^t|D^t)$ by a telescoping loss-change term plus explicit error terms for stochastic gradient noise, quantization variance, local data distribution shift, gradient dissimilarity, and model pruning. The pruning error term is proportional to $\\sum_u \\alpha_u \\delta^t_u \\mathbb{E}[\\|w^t_u\\|^2]$ and does not vanish as learning rates decay unless no weights are removed, so convergence is to a neighborhood of a stationary point rather than to the point itself. The experiments compare 2CEOAFL against the unpruned/unquantized AFL baseline, pruning-only and quantization-only baselines, and centralized SGD; after 100 rounds on CIFAR-10 the proposed algorithm reaches roughly 47.5% test accuracy against about 48.7% for the full AFL baseline, with substantially lower communication overhead.","pith_inferences":["The proof of Theorem 1 applies $\\beta$-smoothness to $f(w^{t+1}|D^{t+1})$ versus $f(w^t|D^t)$, two different loss functions because the dataset changes every round, without bounding the drift of the loss itself; a reader should treat the convergence guarantee as conditional on an unstated loss-drift condition.","A natural stress test is to hold the quantization probability fixed and compare lottery-ticket pruning with random pruning at equal sparsity; the paper's framework predicts that the lottery-ticket mask, not sparsity alone, is what keeps accuracy comparable.","The bound suggests a per-round protocol where the central server adaptively sets $q^t_u$ and $\\delta^t_u$ from radio-resource feedback, because the error terms depend explicitly on both; the authors list joint optimization as future work, so this is an extrapolation."],"forward_implications":["Aerial vehicles can cut per-round computation by training only the pruned subnetwork and cut uplink payload by quantizing gradients, while the global model stays within a few accuracy points of the unpruned, unquantized baseline on the tested CIFAR-10 task.","The convergence bound gives designers a checklist of where accuracy loss comes from: stochastic gradient noise, quantization variance, drifting local data, heterogeneity, and pruning; each term can be read off and, except for pruning, driven down with decaying learning rates.","Because the pruning error term does not vanish unless $\\bar w^{t,0}_u = w^{t,0}_u$, even infinite rounds will not reach a stationary point exactly when any weights are removed; the practical correction is to keep pruning ratios modest.","The trajectory-optimization component connects data freshness to model quality: vehicles fly toward clusters whose time-varying importance is high, so the online dataset and the training objective evolve together rather than in opposition.","The probabilistic quantization rule with probability $q^t_u$ tied to the pruning ratio gives a tuneable knob: when resources are scarce, quantize more; when they are plentiful, send raw gradients."],"supporting_citations":[{"why":"Founded the federated learning setup and aggregation rule that 2CEOAFL modifies for online aerial clients.","marker":"[8]"},{"why":"Supplies the online federated learning formulation and the local data-distribution-shift quantity $\\Phi^t_u$ used in Definition 2 and Theorem 1.","marker":"[9]"},{"why":"Provides the CVXPY solver used to solve the mixed-integer convex trajectory approximation in Algorithm 1.","marker":"[16]"},{"why":"Provides the lottery ticket hypothesis that justifies pruning the dense model to a trainable shallow subnetwork via masks.","marker":"[18]"},{"why":"Defines the low-precision stochastic quantizer of Definition 1 and its unbiased-with-bounded-variance property used in Assumption 4.","marker":"[19]"},{"why":"Supplies the sparsified-SGD pruning-ratio condition (Assumption 5) that ties the mask to the pruning error term.","marker":"[22]"},{"why":"FedPAQ's quantization convergence analysis informs the quantization variance term and the $C_u(q,q^t_u)$ factor.","marker":"[23]"},{"why":"Provides the computation and energy model for pruned federated training used to count time and energy savings.","marker":"[24]"}],"fun_headline_variants":["Pruned, quantized drone federated learning matches full accuracy","Online aerial FL: prune models, compress gradients, keep performance","Drone FL with pruning and quantization rivals full baseline","Aerial federated learning trims models and gradients without losing accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the training objective not changing too much from one round to the next as new data arrives, but the paper only bounds how much the gradients change, never how much the loss itself changes.","fun_headline_variants_meta":{"raw":{"variants":["Pruned, quantized drone federated learning matches full accuracy","Online aerial FL: prune models, compress gradients, keep performance","Drone FL with pruning and quantization rivals full baseline","Aerial federated learning trims models and gradients without losing accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1679,"prompt_tokens":1044,"completion_tokens":635,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":565}},"tokens_in":660,"tokens_out":635,"duration_ms":5918,"temperature":1.0,"reasoning_tokens":565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:11:36.384486+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run 2CEOAFL on a synthetic non-stationary task where the data distribution shifts sharply at a known round, and check whether Theorem 1's inequality (19) holds at that round using measured values; if the expected squared masked-gradient norm exceeds the right-hand side while all stated assumptions hold, the loss-drift step in the proof fails.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Founded the federated learning setup and aggregation rule that 2CEOAFL modifies for online aerial clients."},{"cited_title":"Online-Score-Aided Federated Learning for Resource-Constrained Wireless Clients with Continual Data Arrival","cited_arxiv_id":"2408.05886","evidence_quote":"Supplies the online federated learning formulation and the local data-distribution-shift quantity $\\Phi^t_u$ used in Definition 2 and Theorem 1."},{"cited_title":"CVXPY: A python-embedded modeling language for convex optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the CVXPY solver used to solve the mixed-integer convex trajectory approximation in Algorithm 1."},{"cited_title":"The lottery ticket hypothesis: Finding sparse, trainable neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the lottery ticket hypothesis that justifies pruning the dense model to a trainable shallow subnetwork via masks."},{"cited_title":"QSGD: Communication-efficient SGD via gradient quantization and encoding,","cited_arxiv_id":null,"evidence_quote":"Defines the low-precision stochastic quantizer of Definition 1 and its unbiased-with-bounded-variance property used in Assumption 4."},{"cited_title":"Sparsified SGD with memory,","cited_arxiv_id":null,"evidence_quote":"Supplies the sparsified-SGD pruning-ratio condition (Assumption 5) that ties the mask to the pruning error term."},{"cited_title":"FedPAQ: A communication-efficient federated learning method with periodic averaging and quantization,","cited_arxiv_id":null,"evidence_quote":"FedPAQ's quantization convergence analysis informs the quantization variance term and the $C_u(q,q^t_u)$ factor."},{"cited_title":"Hierarchical federated learning in wireless networks: Pruning tackles bandwidth scarcity and system heterogeneity,","cited_arxiv_id":null,"evidence_quote":"Provides the computation and energy model for pruned federated training used to count time and energy savings."}],"review_version":1}