{"id":"dac9432d-c31f-41b2-9e92-c3e8b42c2240","arxiv_id":"2501.01693","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"DAO-VFL integrates online vertical federated learning with server-side denoising and reinforcement-learning-selected local iteration counts, reporting a regret bound plus experiments on CIFAR-10 and C-MAPSS.","lead":"The authors propose DAO-VFL, a vertical federated learning method for industrial assembly lines that handles streaming sensor data, communication noise, and uneven sensor speeds. It combines a denoising autoencoder on the server with a reinforcement-learning agent that picks how many local updates each sensor runs per round.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Denoising mechanism assumes clean feature embeddings during T_dl, contradicting the paper's noisy-channel premise; without clean data the DAE cannot be trained and the claimed regret improvement collapses.","rationale":"The reader's weakest_assumption identifies exactly this clean-embedding requirement during the denoising learning period, and I agree it is the most load-bearing concern. The paper's system model says all sensor-to-server feature embeddings suffer noise, so the server cannot possess the clean originals needed to train the DAE under Eq. (3). Remark 2 concedes clean data are typically required and explicitly declines to adopt self-supervised alternatives, leaving the contradiction unresolved. Because the entire noise-reduction contribution and the beta_d-dependent regret improvement hinge on this assumption, the central claim is not supported as stated. The missing appendix proof is an additional completeness problem, but the clean-data inconsistency is more fundamental: even a complete proof of Theorem 1 would not rescue the algorithm if the DAE cannot be trained in the claimed setting. The proposed concrete test would settle the concern by removing the clean oracle and measuring whether the denoising benefit survives; if it does not, the REJECT verdict is warranted. No change to the reader's verdict is needed.","tokens_in":19343,"tokens_out":2409,"duration_ms":25385,"concrete_test":"Re-run the C-MAPSS and CIFAR-10 noise experiments without the clean-embedding oracle: during the first T_dl=40 rounds, give the server only the same noisy embeddings it receives later and train the DAE on those noisy pairs (as in Noise2Noise), or disable the DAE entirely and use the NI result as an upper bound. Compare final test RMSE/accuracy and the empirical beta_d against the NE and DAO-NR curves in Fig. 6. If DAO-NR no longer approaches NE (or beta_d is not substantially below beta_n), the T_dl clean-data assumption is load-bearing and the central noise-reduction claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DAO-VFL achieves a tighter regret bound through denoising depends on Assumption 5 (bounded beta_d) and on training the DAE so that beta_d < beta_n. Section IV.2 states: 'During the initial T_dl global rounds, also referred to as the denoising learning period, it is assumed that the original feature embeddings from the sensors are available to the server.' This directly conflicts with the paper's own premise that wireless transmission in industrial environments corrupts all sensor-to-server embeddings with noise. If the channel adds noise to every upload, the server cannot receive clean originals without an unmodeled side channel or redundant clean sensors. Remark 2 acknowledges that clean data are typically required, mentions Noise2Noise/Noise2Void alternatives, but explicitly says these are not explored. Thus the training objective in Eq. (3) is not implementable under the stated system model. Without a trained DAE, beta_d is not reduced relative to beta_n, and the O(T beta_d) term in Theorem 1 remains large; the claimed noise-reduction benefit and the resulting 'tighter regret bound' are unsupported. This is an internal inconsistency in the algorithm's central mechanism, not merely a missing experiment. Separately, the proof of Theorem 1 is deferred to a missing appendix, so the bound itself cannot be independently verified; but the clean-embedding assumption is the more fundamental flaw because it undermines the proposed solution's feasibility.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies online vertical federated learning for multi-sensor industrial assembly lines. It proposes DAO-VFL, in which sensors upload low-dimensional feature embeddings to a server; a denoising autoencoder (DAE) mitigates communication noise, and a deep-reinforcement-learning controller selects per-sensor local iteration counts. The paper derives a regret bound (Theorem 1) under convexity and Lipschitz assumptions, states that effective denoising tightens the bound, and reports experiments on CIFAR-10 and C-MAPSS against noise-included/excluded and homogeneous/heterogeneous baselines. The adaptive-iteration results show latency and reward benefits, and the denoising results show improved accuracy/RMSE relative to the noisy baseline.","tokens_in":19612,"tokens_out":5334,"duration_ms":51724,"significance":"If the theoretical claims were established, the paper would contribute to online VFL in practical IIoT settings by jointly addressing noise and sensor heterogeneity. The problem formulation is relevant and the experimental study is fairly broad. However, the central regret theorem is unverifiable because the proof is not present in the manuscript, its assumptions do not hold in the experiments, and the key denoising benefit is encoded in an assumption rather than derived. The clean-data requirement during the denoising learning period is inconsistent with the paper's own noisy-channel premise. These issues undermine the paper's main claims as submitted. I also note that the manuscript does not ship code or machine-checked proofs, so the theory must stand on the text alone.","major_comments":[{"comment":"The proof of Theorem 1 is deferred to an appendix that is not present in the manuscript; the only indication is the line 'The proof can be found in Appendix.' As a result, the regret bound cannot be independently checked, and the central theoretical contribution is not supported by the submission. A complete proof must be included.","section":"Section V, Theorem 1 and proof"},{"comment":"Theorem 1 relies on Assumption 1, which requires Ft to be convex in Θ. The experiments in Section VII-C use deep convolutional feature extractors (13 conv layers for CIFAR-10 and 2 conv layers for C-MAPSS) followed by nonlinear heads, which are not convex. Therefore the regret guarantee cannot be invoked for the reported empirical results, and the claimed theoretical basis for the observed improvements is absent.","section":"Section V, Assumption 1 vs Section VII-C"},{"comment":"The DAE training objective in Eq. (3) requires paired clean and noisy feature embeddings, and Section IV.2 assumes that 'the original feature embeddings from the sensors are available to the server' during the first Tdl rounds. This contradicts the paper's premise that wireless transmission corrupts all sensor-to-server embeddings. If the channel is noisy for every upload, no clean reference exists; if it is clean for Tdl, the noise model is undefined. Remark 2 acknowledges the issue and mentions Noise2Noise/Noise2Void but states they are not explored. Thus the central noise-reduction mechanism is not implementable under the stated system model.","section":"Section IV.2, Eq. (3) and Remark 2"},{"comment":"Assumption 5 defines βd as the elementwise bound between the denoised gradient and the original gradient. Since the analysis never derives βd from the DAE training in Eq. (3), from the quantization/noise model in Section VII-E, or from any estimation procedure, the statement that 'effective denoising leads to a tighter regret bound' is essentially an assumption, not a conclusion. The comparison with βn is also made without formally defining βn. To support the paper's central claim, βd must be derived or measured under a concrete noise model, with a demonstrated reduction relative to βn.","section":"Section V, Assumption 5 and subsequent findings"}],"minor_comments":[{"comment":"The tuple lists Υco_t,k twice; the second entry should be the communication latency Υcm_t,k.","section":"Algorithm 2, line 4"},{"comment":"Eq. (20) is called a 'gain function' but it is a squared temporal-difference loss to be minimized by the critic; the description of updating the critic by 'maximizing the reward via Eq. 20' is confusing.","section":"Eq. (20) and Algorithm 2"},{"comment":"The benchmark 'DAO-NR' appears in figures and text (e.g., Fig. 6) without being defined; the reader must infer that it is the noise-reduction variant of DAO-VFL.","section":"Section VII-F"},{"comment":"Theorem 1's left-hand side uses E_t[Ft(...)] while Eq. (2) defines regret without an expectation; the meaning of E_t is not explained.","section":"Theorem 1 and Eq. (2)"},{"comment":"The parameters µ0 and µ in the collection latency definition, Eq. (10), are not defined before use; they are only assigned example values in Section VII-D.","section":"Section IV.2 and Section VII-D"}],"recommendation":"reject","confidential_remarks":"The manuscript would require major restructuring and a complete proof before resubmission. I do not see a path to acceptance without resolving the clean-data inconsistency and the circularity of βd; the missing appendix alone would also block review. I recommend reject rather than major revision because the central mechanism depends on assumptions that are not established in the current text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible new combination—online vertical federated learning, a server-side denoising autoencoder, and DRL-chosen local iteration counts—packaged for the IIoT assembly-line scenario. The latency model and the fairness analysis are reasonable, and the experiments on C-MAPSS and CIFAR-10 give some evidence that adaptive local iterations cut latency. But the paper's center does not hold as written. The proof of Theorem 1 is deferred to an appendix that the arXiv v1 doesn't contain, so the regret bound is unverifiable. And the bound's key message—that effective denoising tightens regret—is essentially baked into Assumption 5, which defines beta_d as the difference between denoised and original gradients. That's a tautology unless you can show the DAE actually makes beta_d small, and the paper doesn't derive that from the training procedure.\n\nThe bigger problem is internal consistency. The whole motivation is that wireless uploads corrupt feature embeddings with noise. But Section IV.2 assumes that during the initial T_dl rounds, the server receives the original clean embeddings to train the DAE. Remark 2 acknowledges that clean data are needed and mentions Noise2Noise/Noise2Void as alternatives, but explicitly doesn't explore them. So Eq. (3)'s training objective is not implementable under the stated system model. Without a trained DAE, the O(T beta_d) term in Theorem 1 just stays large, and the claimed noise-reduction benefit is unsupported.\n\nThere's also an evaluation problem. The DRL agent's reward uses test accuracy (Algorithm 2, line 9), so the reported test accuracy is contaminated by test-set-in-the-loop. The paper does honestly show DAO-PPO doesn't beat the homogeneous baseline on test accuracy (Fig. 8), and the abstract's \"superior performance\" is overstated—the actual advantages are latency and reward, not accuracy. There are also no comparisons against existing online VFL baselines like [38]; the benchmarks are all ablations of the proposed method.\n\nWhat's worth keeping: the problem formulation is clear, the latency/communication model is reasonable, and the fairness insight—that balanced local iteration counts tighten the bound—is supported by a controlled comparison. The paper is not sloppy in its internal logic; it just overreaches.\n\nWho this is for: researchers working on VFL for IIoT will find the framing useful, but the theory and the central denoising claim need major work. I'd send it to peer review, but with a firm request for the appendix proof, a fix for the clean-embedding assumption (or a realistic training protocol), and a re-evaluation using a held-out test set. As it stands, it's not publishable.","headline":"A plausible new combination of online VFL, denoising, and DRL-based local iterations, but the proof is missing and the denoising mechanism assumes the very clean data the paper says isn't available.","tokens_in":20158,"tokens_out":3481,"would_cite":false,"duration_ms":32877,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","68W27"],"pacs":[],"model":"deepseek-v4-flash","headline":"Denoising plus adaptive local steps tightens the regret bound for online vertical federated learning in noisy industrial settings.","keywords":["Industrial Internet of Things","Vertical Federated Learning","Online Learning","Deep Reinforcement Learning","denoising autoencoder","regret bound","adaptive local iterations","multi-sensor data"],"falsifier":"Run DAO-VFL in a setting where no clean feature embeddings ever reach the server, so the autoencoder can only be trained on noisy pairs, and compare cumulative regret against the paper's noise-included and noise-excluded baselines; if the denoised run does not beat the noise-included run, or if empirical regret grows faster than $O(\\sqrt{T}+T\\beta_d)$, the central claim is falsified.","tokens_in":19084,"feed_emoji":"🏭","tokens_out":11353,"duration_ms":101104,"temperature":0.7,"pith_summary":"DAO-VFL targets an industrial assembly line where multiple sensors with distinct feature spaces train a shared model online by uploading feature embeddings to a server over noisy wireless links. The paper's central claim is that a server-side denoising autoencoder plus per-sensor adaptive local iteration counts makes this online problem learnable: under convexity and bounded-gradient assumptions, cumulative regret grows as $O(\\sqrt{T}+T\\beta_d)$, where $\\beta_d$ measures how close the denoised gradient is to the clean gradient. Because $\\beta_d$ sits inside the regret bound, effective denoising directly buys a tighter guarantee, and keeping local iteration counts similar across sensors tightens the bound as well. If the claim holds, vertical federated learning can run on streaming industrial data despite transmission noise and heterogeneous sensor capabilities.","feed_headline":"Denoising plus adaptive local steps tightens federated regret","feed_subtitle":"A server-side denoiser plus per-sensor adaptive local updates shrinks the regret term controlled by noise.","key_machinery":"The load-bearing mechanism is the server-side denoising autoencoder applied to noisy feature embeddings, paired with a deep reinforcement learning agent that outputs the local iteration count $E_{t,k}$ for each sensor in every global round. The denoising autoencoder is trained during an initial denoising learning period to map the noisy embedding $\\tilde{h}_k(\\theta_k;\\mathbf{x}_k)$ to a denoised $\\hat{h}_k(\\theta_k;\\mathbf{x}_k)$, which is what makes the gradient-deviation bound $\\beta_d$ in Assumption 5 small. The server bundles the head model and all denoised embeddings into the model representation $\\hat{\\Phi}^{t,0}$, distributes it to sensors, and each sensor runs online gradient descent for $E_{t,k}$ iterations. The regret analysis then separates the denoising error $\\beta_d$ from the iteration-count terms $E_{\\max}$ and $E_{\\min}$, so the theorem's tightness depends on exactly the two components the algorithm controls.","core_discovery":"On the paper's own terms, the discovery is that online vertical federated learning remains tractable in noisy, heterogeneous industrial settings when the server denoises incoming feature embeddings and each sensor's local iteration count is chosen adaptively. Theorem 1 gives, under Assumptions 1 to 5, the regret bound $$\\mathrm{Reg}_T \\le \\frac{\\|\\$Theta^{{1,0}}$-\\Theta^*\\|^2}{2\\eta E_{\\min}} + \\frac{\\eta T D \\$beta_d^{2}$}{E_{\\min}} + \\frac{\\eta T E_{\\max} $L^{2}$ K}{E_{\\min}} + 2DT\\rho(\\eta\\$\\lambda$ E_{\\max}L + \\beta_d).$$ With $\\eta = O(1/\\sqrt{T})$, this is $O(\\sqrt{T}+T\\beta_d)$, where $\\beta_d$ is the per-coordinate bound on the difference between denoised and clean gradients from Assumption 5. The paper argues that replacing $\\beta_d$ with the larger noise-level bound $\\beta_n$ exposes why denoising helps, and that small $E_{\\max}$ with large $E_{\\min}$ tightens the bound. Experiments on CIFAR-10 and C-MAPSS are presented as evidence that the denoised variant matches or improves on the noise-free baseline, and that the DRL-based iteration choices reduce total latency and improve the reward objective.","pith_inferences":["The paper leaves implicit that the clean-embedding training period could be replaced by self-supervised denoising trained only on noisy pairs; a testable extension is to analyze whether the regret bound still holds with $\\beta_d$ defined against such a denoiser.","Because the regret term $O(T\\beta_d)$ is additive, there is a point of diminishing returns where further denoising effort stops mattering; measuring $\\beta_d$ empirically would let practitioners stop investing in denoising once other terms dominate.","The adaptive iteration mechanism is effectively a straggler-mitigation policy: sensors with slow CPUs or poor channels receive fewer local iterations, which connects DAO-VFL to asynchronous and heterogeneous federated learning beyond the assembly-line setting.","A direct experimental check would measure the empirical $\\beta_d$ of the trained autoencoder and compare the observed regret against the bound's prediction, turning Theorem 1 into a deployable diagnostic."],"forward_implications":["A practitioner can tune the denoiser and the local iteration schedule by targeting $\\beta_d$, $E_{\\max}$, and $E_{\\min}$ instead of tuning the whole system empirically.","Partially effective denoising still improves the guarantee, because any $\\beta_d$ smaller than the raw noise bound $\\beta_n$ makes the regret bound tighter.","Keeping per-sensor local iteration counts close to each other is not only a latency fairness goal; the regret bound makes it a formal requirement.","The algorithm can be applied to streaming IIoT data without waiting for a static dataset, since both feature and head models update by online gradient steps.","The reward design of the DRL problem gives a concrete trade-off among accuracy, total latency, and iteration disparity, and the learned policy executes that trade-off each round."],"supporting_citations":[{"why":"Establishes the online vertical federated learning regret framework and notation that DAO-VFL extends with denoising and adaptive local iterations.","marker":"[38]"},{"why":"Supplies Assumptions 1-4 (convexity, Lipschitz gradients, bounded models) that Theorem 1 relies on.","marker":"[44]"},{"why":"Gives the online gradient descent update rule used for local feature and head model updates.","marker":"[40]"},{"why":"Provides the proximal policy optimization algorithm used to train the DRL agent that chooses local iteration counts.","marker":"[47]"},{"why":"Provides the C-MAPSS turbofan degradation dataset used for the remaining-useful-life experiments.","marker":"[48]"},{"why":"Cited as a clean-data-free denoising alternative that could replace the clean-embedding training period; the paper does not analyze this extension.","marker":"[42]"},{"why":"Same role as [42]: a single-noisy-image denoising method that could train the DAE without clean targets.","marker":"[43]"}],"fun_headline_variants":["Denoising tightens regret in online vertical federated learning","Adaptive local steps curb noise in federated IoT learning","Denoising shrinks federated regret in noisy industrial data","Adaptive denoising curbs regret in vertical federated learning","Online federated learning made robust to sensor noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Section IV.2 assumes that during the first $T_{dl}$ global rounds the server receives clean, noise-free feature embeddings to train the denoising autoencoder, and Remark 2 concedes that clean data is typically required; since the paper's motivating scenario says all sensor-to-server wireless transmissions are noisy, this clean-access period is the load-bearing premise.","fun_headline_variants_meta":{"raw":{"variants":["Denoising tightens regret in online vertical federated learning","Adaptive local steps curb noise in federated IoT learning","Denoising shrinks federated regret in noisy industrial data","Adaptive denoising curbs regret in vertical federated learning","Online federated learning made robust to sensor noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2278,"prompt_tokens":1039,"completion_tokens":1239,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":1156}},"tokens_in":655,"tokens_out":1239,"duration_ms":8657,"temperature":1.0,"reasoning_tokens":1156,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:21:53.253096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DAO-VFL in a setting where no clean feature embeddings ever reach the server, so the autoencoder can only be trained on noisy pairs, and compare cumulative regret against the paper's noise-included and noise-excluded baselines; if the denoised run does not beat the noise-included run, or if empirical regret grows faster than $O(\\sqrt{T}+T\\beta_d)$, the central claim is falsified.","supporting_citations":[{"cited_title":"Online vertical federated learning for cooper- ative spectrum sensing,","cited_arxiv_id":null,"evidence_quote":"Establishes the online vertical federated learning regret framework and notation that DAO-VFL extends with denoising and adaptive local iterations."},{"cited_title":"Fedqogd: Federated quantized online gradient descent with distributed time-series data,","cited_arxiv_id":null,"evidence_quote":"Supplies Assumptions 1-4 (convexity, Lipschitz gradients, bounded models) that Theorem 1 relies on."},{"cited_title":"Online gradient descent learning algorithms,","cited_arxiv_id":null,"evidence_quote":"Gives the online gradient descent update rule used for local feature and head model updates."},{"cited_title":"Damage propaga- tion modeling for aircraft engine run-to-failure simulation,","cited_arxiv_id":null,"evidence_quote":"Provides the C-MAPSS turbofan degradation dataset used for the remaining-useful-life experiments."},{"cited_title":"Noise2noise: learning image restoration without clean data,","cited_arxiv_id":null,"evidence_quote":"Cited as a clean-data-free denoising alternative that could replace the clean-embedding training period; the paper does not analyze this extension."},{"cited_title":"Noise2void-learning denoising from single noisy images,","cited_arxiv_id":null,"evidence_quote":"Same role as [42]: a single-noisy-image denoising method that could train the DAE without clean targets."}],"review_version":1}