{"id":"33d81cbf-0584-4344-8f80-97a88f760076","arxiv_id":"2505.11648","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A joint graph-learning and signal-restoration aggregation step, solved as a difference-of-convex program, improves federated learning accuracy under noisy client updates.","lead":"This paper addresses what happens when the data used to update an AI model gets corrupted while being sent from devices to a central server. It lets the server learn which devices behave similarly and use that map to repair noisy or missing updates, improving accuracy by a few percentage points in tests.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's gradient formulas do not match the DC objective in Eq. (14), so the published PDCA update is not a valid descent step; the convergence claim and reproducibility are undermined.","rationale":"The reader's weakest_assumption is the graph-smoothness assumption of Assumption 1. My concern is different and more fundamental: the algorithm as written is internally inconsistent with the DC objective it claims to minimize. The printed gradient formulas in Section 4.2 use the adjoint operator D* on a non-symmetric matrix and with incorrect scaling, so the PDCA update is not a valid descent step for Eq. (14). This is a concrete, checkable mathematical error, not a modeling assumption. However, it does not by itself falsify the empirical accuracy claim in Table 1, because the experiments may have used corrected gradients not faithfully reported. Therefore the correct verdict remains CONDITIONAL, with a new high-priority condition: the authors must verify or correct the PDCA gradients, ideally by providing code or a finite-difference check. The reader's verdict is unchanged, but the conditions are sharpened. My agreement_with_reader is 'disagree' because the reader identified Assumption 1 as the weakest point, whereas I find the gradient inconsistency more load-bearing for the paper's methodological claim. The proposed concrete_test would settle whether the concern lands; if it does, the paper's theoretical justification collapses even though the empirical numbers might survive as a heuristic.","tokens_in":9489,"tokens_out":22778,"duration_ms":208567,"concrete_test":"For K=2,d=1, set α=1, w=1, Ψ=(0,1)^T, and zero μ/other terms. Compute the finite-difference gradient of f(Ψ,w)=α||w+Tvec(D(Ψ))||^2 at this point; it is (-8,8). Then evaluate Algorithm 1's ∇Ψf second term with the same inputs (it returns (-4,2)). Repeat at K=3 with random Ψ,w and compare with automatic differentiation using PyTorch/JAX. If the norms of the differences exceed 1e-4, the printed gradients are incorrect, and the authors must supply corrected formulas or code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2's PDCA update is not a gradient step for the stated DC objective (14). Specifically, ∇Ψf is given as αD*(vec^{-1}(T^T(w+Tvec(D(Ψ)))))Ψ and ∇Ψh as αD*(vec^{-1}(T^TTvec(D(Ψ))))Ψ, with D*(H)=diag(H1)+diag(H^T1)-2H. Both formulas apply D* to a non-symmetric H whose lower triangle is zero because T^T only back-projects the upper triangle. For the simplest case K=2, α=1, w=1, Ψ=(0,1), the finite-difference derivative of α||w+Tvec(D)||^2 is (-8,+8), while the printed ∇Ψf term evaluates to (-4,2); using the symmetrized H=(W+D) would give (-4,+4) and still miss the factor 2. The h-gradient has the same structural error. Hence Algorithm 1 does not descend on (14), the claimed local convergence guarantee has no basis, and a reader implementing the text will not reproduce Table 1. This is more load-bearing than Assumption 1 because it breaks the link between the method and its implementation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a robust federated-learning aggregation method that jointly estimates an inter-client graph and restores noisy/missing model parameters during the global model update. The aggregation problem is formulated as a difference-of-convex (DC) optimization, solved by a proximal DC algorithm (PDCA), and evaluated on MNIST and CIFAR-10 with five published baselines. The central claims are that the joint formulation improves classification accuracy by 2–5% under biased distributions and transmission noise, and that the DC/PDCA formulation provides convergence guarantees.","tokens_in":9747,"tokens_out":13306,"duration_ms":134578,"significance":"If the algorithmic issues are repaired, the paper offers a clean and useful idea: unifying graph learning and signal restoration inside the FL aggregation step, with a DC decomposition that is algebraically sound in Eq. (13). The reported accuracy gains against five published baselines are concrete and are evaluated on external datasets, so the empirical claims are not circular. The paper also includes a two-step baseline, which helps isolate the benefit of joint optimization. Counterbalancing these strengths, the manuscript does not provide code, proof sketches, or sensitivity analyses, and the experimental setup is small (20 clients, 30 rounds, simple CNN). The main concern is that the printed gradient formulas in Section 4.2 do not match the stated objective, which undermines the reproducibility of the algorithm and the convergence claim as written.","major_comments":[{"comment":"The printed gradients ∇Ψf and ∇Ψh are not the derivatives of f and h defined in (15). For K=2, α=1, w=1, and Ψ=(0,1), the true gradient of α||w+Tvec(D(Ψ))||² is (-8,8), whereas the printed formula αD*(vec^{-1}(T^T(w+Tvec(D(Ψ)))))Ψ evaluates to (-4,0). Even if the argument of D* were symmetrized, the result would be (-4,4), still off by a factor of 2. The h-gradient has the same structural problem: the true derivative of α||Tvec(D(Ψ))||² is (-4,4), but the printed expression gives (-2,0). Consequently, the vector u-s used in Algorithm 1 is not the gradient of f-h, the update is not a PDCA step for (14), and the claimed convergence guarantee is unsupported. Please correct the gradient formulas (the correct forms involve 4α(diag(V)-V)Ψ with V=W+D(Ψ) for f, and 4α(diag(D)-D)Ψ for h) and re-verify the experiments.","section":"Section 4.2, Algorithm 1 and Eq. (15)"},{"comment":"The mapping from (11) to (14) is not an exact equivalence as claimed. In (11), the regularizer is γ||W||₁ over a symmetric, zero-diagonal W, which equals 2γ||w||₁ for w=upper(W). Equation (15b) instead uses γ||w||₁. This changes the objective by a factor of 2 unless γ is implicitly redefined; the paper does not state this. Please correct the factor or explicitly absorb it into the hyperparameter.","section":"Section 4.1, Eq. (15b)"},{"comment":"The convergence claim is not supported. The footnote states that the PDCA conditions are satisfied but omits the details, and the paper does not identify which conditions are needed for the proximal DC algorithm or verify them for the specific f, g, and h in (15). Moreover, because the gradients in Algorithm 1 are incorrect, the update is not a valid PDCA iteration even if the abstract conditions were met. Please either state and verify the conditions or substantially weaken the convergence statement in the Introduction.","section":"Section 4.2, footnote 2"},{"comment":"Assumption 1—that the noisy/missing client updates vary smoothly on the inter-client graph—is the modeling premise of the entire aggregation objective, but it is never validated. No experiment reports the quality of the learned graph (e.g., edge recovery against a known graph, smoothness of the restored Ψ, or behavior of the learned graph as noise increases). Without such evidence, the robustness claim rests on an unexamined assumption, although this is distinct from the accuracy comparison in Table 1.","section":"Section 3.2 and Section 5"}],"minor_comments":[{"comment":"The definition of the Euclidean distance matrix contains an erroneous summation symbol: [D(X)]mn should be ||x_m - x_n||², not Σ_{m,n}||x_m - x_n||².","section":"Section 3.3"},{"comment":"The initialization of w(r+1,0) is said to use cosine similarity among the row entries of the received parameters, but the optimization objective (11) is built on squared Euclidean distances. Please specify the exact initialization used, since different initial graphs can affect the local minimum reached.","section":"Algorithm 1 and Section 5.1"},{"comment":"The paper reports 10 independent runs but does not report standard deviations or error bars. The 2–5% accuracy differences in Table 1 would be more convincing with variance information, and Fig. 4 would be stronger if at least FedAvg and FedAMP were included for comparison.","section":"Table 1 and Fig. 4"},{"comment":"The hyperparameters α=0.05 and μ=β=γ=ρ=1.0 are described as chosen experimentally, but no sensitivity analysis is provided. It is therefore unclear whether the reported gains are robust to hyperparameter changes or are a result of fine-tuning.","section":"Section 5.1"},{"comment":"Reference [23] is cited for CIFAR-10 but appears to point to the AlexNet paper; the usual citation is Krizhevsky's 'Learning Multiple Layers of Features from Tiny Images' (2009).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper reads like a compressed workshop paper, and many supporting details are omitted. The gradient error in Section 4.2 is the key technical problem; it is likely fixable by a correct derivation and by re-running the experiments to confirm that Table 1 was produced with the corrected update. I would encourage the authors to provide the corrected formulas, state the PDCA conditions explicitly, and report the factor issue in the L1 norm term. If the empirical results survive these corrections, the paper could be acceptable after a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the idea is sensible and the writing is mostly clear, but there is a load-bearing error in the algorithm section. The joint formulation of graph learning and signal restoration inside the FL aggregation step is a legitimate extension of existing graph-based FL (CFL, FedAMP, pFedGraph), which build the inter-client graph first and then aggregate. The DC decomposition in (13)–(15) is algebraically correct, and the experimental setup is honest: five published baselines, two datasets, non-IID partitions via Dirichlet, and a two-step ablation. For a workshop paper, that is decent.\n\nThe soft spots: Section 4.2's gradients do not match (14). For K=2, Ψ=(0,1), w=1, the true gradient of the α||w+Tvec(D(Ψ))||² term is (−8,+8), while the printed ∇Ψf evaluates to (−4,2); the h-gradient has the same structural error. Algorithm 1 is not a valid PDCA step on the stated objective. The footnote saying convergence conditions are met is hand-waved, and the convergence claim has no basis for the algorithm as written. A reader implementing the text will not reproduce Table 1. This is not a minor typo; it breaks the link between the method and its implementation.\n\nOther issues, more minor: the missing-rate experiment has no baseline; confidence intervals are absent; no code or data artifacts are provided; Assumption 1 (smoothness) is not verified under noise. The gains are modest (2–5 points on MNIST/CIFAR-10), so the method does not enable a qualitatively new capability. The hyperparameters are manually set. The citation pattern is fine.\n\nCredit where due: the DC decomposition is the right kind of move for the biconvex problem, and the comparison set is reasonable. The central accuracy claim is consistent with the table as far as it goes, but because the algorithm's gradients are broken, I cannot treat the table as evidence for the method as written.\n\nThis paper is for readers interested in graph-signal-processing approaches to robust FL; they might get ideas from the formulation, but they should not implement from this text. A serious editor should send it to peer review, but the review would rightly demand corrected gradients, a derivation or reference for them, and either code or a missing-rate baseline. My own verdict is reject and resubmit.","headline":"A sensible joint graph-learning and signal-restoration formulation whose published algorithm has wrong gradients, so the experiments cannot be reproduced from the text.","tokens_in":10256,"tokens_out":6578,"would_cite":false,"duration_ms":59498,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Jointly learning a client graph while restoring noisy model parameters gives federated learning a 2–5% accuracy edge in biased, noisy settings.","keywords":["federated learning","graph learning","signal restoration","difference-of-convex optimization","proximal DC algorithm","robust aggregation","graph signal processing","noisy communication"],"falsifier":"Run the method on a synthetic federated task with known client clusters and randomly permute client identities each round so the parameters are not smooth on any fixed graph; if the accuracy margin over FedAvg persists, the graph-smoothness assumption is not what produces the gain.","tokens_in":9302,"feed_emoji":"🕸️","tokens_out":6979,"duration_ms":64707,"temperature":0.7,"pith_summary":"Federated learning sends model parameters from clients to a server, and those parameters are often corrupted by noise or missing entries when bandwidth is low. This paper claims that the server can aggregate more robustly by solving one joint problem instead of two separate steps: learn a graph encoding which clients' parameter vectors are similar, and restore a clean version of the parameter matrix from the noisy received one. The two tasks are coupled through a graph-smoothness assumption—restored parameters should vary smoothly across the learned client graph—and the whole objective is rewritten as a difference-of-convex program solved by a proximal DC algorithm with convergence to a local optimum. On MNIST and CIFAR-10 with biased data splits and additive Gaussian noise, the proposed aggregation outperforms standard and graph-based federated baselines by up to 2–5% in classification accuracy, and it stays within 10% of its clean-signal accuracy when up to 10% of parameters are missing.","feed_headline":"Joint graph learning restores noisy federated updates","feed_subtitle":"A server that learns client similarities while cleaning model parameters gains 2–5% accuracy on biased, noisy FL.","key_machinery":"The load-bearing mechanism is the difference-of-convex decomposition of the biconvex smoothness coupling. In the joint objective, the term $\\operatorname{tr}(W D(\\Psi))$ is convex in $W$ for fixed $\\Psi$ and convex in $\\Psi$ for fixed $W$, but not jointly convex. The identity $\\operatorname{tr}(W D(\\Psi)) = \\tfrac{1}{2}\\|W + D(\\Psi)\\|_F^2 - \\tfrac{1}{2}(\\|W\\|_F^2 + \\|D(\\Psi)\\|_F^2)$ rewrites it as a convex term minus a convex term, giving a DC program with components $f$, $g$, and $h$. The proximal DC algorithm then alternates a gradient step on the difference and a proximal step on $g$, producing a local minimizer with finite-iteration convergence guarantees. The graph Laplacian smoothness assumption provides the semantic content: it is what lets the learned $W$ act as a low-pass filter that restores missing or noisy parameter entries from their neighbors.","core_discovery":"The paper's central claim is that graph-based aggregation in federated learning should not treat the inter-client graph as a fixed input. Under Assumption 1, the aggregated parameter matrix $\\Psi$ is assumed to vary smoothly over an unknown client graph, so the server should minimize a Laplacian smoothness term $\\operatorname{tr}(\\Psi^\\top L\\Psi)$ while simultaneously reconstructing $\\Psi$ from the noisy, partially masked received matrix $\\tilde X_{\\mathrm{local}}$. The authors fold the smooth graph-learning objective and the graph-filtered aggregation into one biconvex problem, then decompose the biconvex term $\\operatorname{tr}(W D(\\Psi))$ as a difference of two convex functions using the identity $\\|W + D(\\Psi)\\|_F^2 - (\\|W\\|_F^2 + \\|D(\\Psi)\\|_F^2)$. Solving that DC program with the proximal DC algorithm yields a sequence that converges to a local optimum. The reported result is that this joint estimator beats two-step graph-learning-then-aggregation and the FedAvg, FedProx, CFL, FedAMP, and pFedGraph baselines on MNIST and CIFAR-10 at noise levels 0.1 and 0.2, with the largest gap being 95.17% versus 93.12% on MNIST at noise level 0.1.","pith_inferences":["Inference: the per-round learned inter-client graph could be tracked across communication rounds as a privacy-preserving diagnostic for client drift or clustering, since it is derived only from transmitted parameters, not raw data.","Inference: the same DC decomposition applies to any bilinear smoothness term between a signal and a graph, so the method should transfer to restoring gradients, activations, or sensor-network signals under missing entries, not just federated model parameters.","Inference: because the DC formulation has a per-round cost that grows with the number of client pairs, practical large-scale deployment would likely need a sparse or block approximation of the learned graph; the paper does not address this scaling.","Inference: a natural test of the smoothness assumption would be to compare the learned graph with a ground-truth similarity graph computed from clean parameters before adding noise, reporting edge-recovery quality alongside accuracy."],"forward_implications":["Under biased data distributions and Gaussian transmission noise at levels 0.1 and 0.2, joint graph estimation and restoration yields higher average test accuracy than all five compared baselines on both MNIST and CIFAR-10.","On MNIST at noise level 0.1 the proposed method reaches 95.17% accuracy versus 93.12% for the best baseline FedAMP, while the two-step alternating solver reaches only 88.99%, showing that joint optimization is the source of the gain.","With missing rates from 0 to 0.1, the method keeps accuracy within 10% of its noise-free level on both datasets, indicating it can compensate for absent parameters using the learned graph.","Because the graph-filtered aggregation specializes to cluster-wise and adjacency-wise graph aggregation in limiting cases, the proposed objective extends existing graph-based federated aggregations rather than replacing them."],"supporting_citations":[{"why":"Supplies FedAvg model averaging as the standard baseline and weights clients by dataset size; the proposed method must beat it.","marker":"[7]"},{"why":"Defines the FedAMP adjacency-wise graph-based aggregation baseline.","marker":"[8]"},{"why":"Defines pFedGraph, an adjacency-wise graph-based aggregation baseline with inferred collaboration graphs.","marker":"[9]"},{"why":"Introduces graph-signal smoothness and the Laplacian quadratic form that motivates Assumption 1.","marker":"[11]"},{"why":"Provides the proximal DC algorithm with convergence guarantees used to solve the reformulated problem.","marker":"[12]"},{"why":"Defines clustered federated learning (CFL), the cluster-wise graph-aggregation baseline.","marker":"[15]"},{"why":"Supplies the smooth graph-learning objective with log penalty and sparsity that the joint problem builds on.","marker":"[18]"},{"why":"Models federated learning over wireless fading channels and noisy parameter reception, justifying the noise model in Eq. (10).","marker":"[20]"},{"why":"Models over-the-air federated learning with noisy aggregated parameters, a second source for the degradation model.","marker":"[21]"},{"why":"Provides the alternating minimization (two-step) solver used as the decomposed baseline for the joint problem.","marker":"[24]"}],"fun_headline_variants":["Joint graph learning fixes noisy federated updates","Cleaner federated models via learned client graphs","Graph-aware aggregation beats noise in federated learning","Restore noisy FL parameters with joint graph estimation","Joint graph estimation lifts noisy FL accuracy by up to 5%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true model parameters from different clients differ smoothly from one another according to an unknown similarity graph; if the noisy updates do not have that smooth structure, jointly learning the graph and restoring the parameters could amplify the noise instead of removing it.","fun_headline_variants_meta":{"raw":{"variants":["Joint graph learning fixes noisy federated updates","Cleaner federated models via learned client graphs","Graph-aware aggregation beats noise in federated learning","Restore noisy FL parameters with joint graph estimation","Joint graph estimation lifts noisy FL accuracy by up to 5%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000436,"raw_usage":{"total_tokens":2235,"prompt_tokens":981,"completion_tokens":1254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1180}},"tokens_in":597,"tokens_out":1254,"duration_ms":9547,"temperature":1.0,"reasoning_tokens":1180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:52:10.874327+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a synthetic federated task with known client clusters and randomly permute client identities each round so the parameters are not smooth on any fixed graph; if the accuracy margin over FedAvg persists, the graph-smoothness assumption is not what produces the gain.","supporting_citations":[{"cited_title":"First, we assume that the model param- eters, aggregated in the global model update, smoothly vary on the inter-client graph","cited_arxiv_id":null,"evidence_quote":"Supplies FedAvg model averaging as the standard baseline and weights clients by dataset size; the proposed method must beat it."},{"cited_title":"Advances and open problems in federated learn- ing,","cited_arxiv_id":null,"evidence_quote":"Defines the FedAMP adjacency-wise graph-based aggregation baseline."},{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Defines pFedGraph, an adjacency-wise graph-based aggregation baseline with inferred collaboration graphs."},{"cited_title":"Robust federated learning with noisy communication,","cited_arxiv_id":null,"evidence_quote":"Provides the proximal DC algorithm with convergence guarantees used to solve the reformulated problem."},{"cited_title":"Personalized cross-silo federated learning on non- iid data,","cited_arxiv_id":null,"evidence_quote":"Defines clustered federated learning (CFL), the cluster-wise graph-aggregation baseline."},{"cited_title":"The emerging field of signal processing on graphs: Extending high-dimensional data analysis to net- works and other irregular domains,","cited_arxiv_id":null,"evidence_quote":"Supplies the smooth graph-learning objective with log penalty and sparsity that the joint problem builds on."},{"cited_title":"Federated optimization in heterogeneous networks,","cited_arxiv_id":null,"evidence_quote":"Models federated learning over wireless fading channels and noisy parameter reception, justifying the noise model in Eq. (10)."},{"cited_title":"SCAFFOLD: Stochastic controlled averag- ing for federated learning,","cited_arxiv_id":null,"evidence_quote":"Models over-the-air federated learning with noisy aggregated parameters, a second source for the degradation model."},{"cited_title":"Fed- erated multi-task learning,","cited_arxiv_id":null,"evidence_quote":"Provides the alternating minimization (two-step) solver used as the decomposed baseline for the joint problem."}],"review_version":1}