{"id":"bdd5202c-9547-4113-a8b5-309cdc427dcc","arxiv_id":"2506.10207","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"FedMLAC couples personalized local audio models with a shared plug-in model via bidirectional knowledge distillation, plus layer-wise pruning aggregation, to jointly address data, model, and label heterogeneity in federated audio classification.","lead":"A new federated learning framework for audio classification combines mutual learning between local and global models with a layer-wise pruning aggregation to handle uneven data, different model sizes, and corrupted labels or noise. The method reports accuracy gains of roughly 1 to 5 percent over existing federated baselines across four audio benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(1/sqrt(T)) convergence theorem is not established: it assumes LPA's pruning bound (Assumption 5) and jumps from per-client local losses to a bound on the global gradient without a comparison argument; the plug-in global objective is never defined.","rationale":"The paper is a plausible engineering contribution: it combines mutual learning with trimmed layer-wise aggregation, and the experiments cover four benchmarks, several heterogeneity levels, and noise conditions. The reader's conditional verdict is appropriate. My stress test focuses on the theoretical claim because it is listed as a contribution and is presented as formal evidence for the same robustness the paper claims empirically. The weakest point is not merely that Assumption 5 is strong; it is that the proof does not connect the global plug-in sequence to the global objective. Eq. 27's descent is in terms of local losses; Eq. 28's conclusion is in terms of nabla L(theta^g_t). In standard FedAvg analyses this gap is bridged by defining a virtual global sequence or bounding client drift; no such argument appears here. Moreover, since the plug-in is trained only with Eq. 9's KL loss, the 'global loss L(theta^g)' used in Assumption 6 is undefined in the methodology. These are internal issues, not just disagreement with consensus. I would keep the reader's CONDITIONAL verdict: the empirical results may survive, but the theoretical claim needs either a repaired proof or an explicit weakened statement. If the authors provide the missing derivation with explicit delta^2 behavior, the concern is resolved.","tokens_in":20051,"tokens_out":7782,"duration_ms":105278,"concrete_test":"Independently re-derive Theorem 1 from Algorithms 1-2, defining L(theta^g) explicitly and replacing Assumption 5 by a bound derived from Eq. 12's pruning rule. Track theta^g_t through Eq. 14 and show where ||nabla L(theta^g_t)||^2 appears; if the only available decrease is over per-client local gradients with no global comparison, the theorem is unsupported. As an empirical cross-check, log the trusted-set deviations delta^l_t on the GSC label-error runs; if they grow with the label error rate or with T, Assumption 5 is violated in the regime the paper claims to analyze.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4's Theorem 1 (Eq. 28) is the paper's theoretical support for the central robustness claim, but it is not actually derived from the algorithm. First, Assumption 5 postulates E[||theta^{k,l}_{t+1} - bar theta^l_{t+1}||^2] <= delta^2 for the trusted set. LPA only removes the largest/smallest deviations from an average that may itself be corrupted; it gives no bound on the remaining deviations, so delta^2 is not established. Second, even granting Assumption 5, Eq. 27 bounds a weighted sum of per-client local-loss decreases, whereas Eq. 28 claims a bound on (1/T) sum E||nabla L(theta^g_t)||^2 for the global plug-in objective. No comparison step (e.g., a virtual global sequence or a bound on ||nabla L(theta^g_t)|| in terms of the trusted local gradients) is supplied, and the global objective L(theta^g) minimized by the KL-only update in Eq. 9 is never defined. Without one of these steps, the O(1/sqrt(T)) rate is not a theorem about FedMLAC as specified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedMLAC, a federated learning framework for audio classification that aims to jointly address data heterogeneity, model heterogeneity, and data poisoning. Each client trains a personalized local model together with a lightweight, globally shared Plug-in model, using bidirectional KL-divergence-based mutual learning. The server aggregates Plug-in updates with a Layer-wise Pruning Aggregation (LPA) that removes per-layer outliers before weighted averaging. The authors claim an O(1/sqrt(T)) convergence rate for the global Plug-in model and report experiments on GSC, IEMOCAP, CREMA-D, and Urban Sound under homogeneous and heterogeneous model settings, with robustness tests under additive noise and label errors.","tokens_in":20330,"tokens_out":4373,"duration_ms":46628,"significance":"If the framework works as described, it would be a useful unified contribution to federated audio classification, since most prior work addresses only one or two of the three challenges. The empirical study is fairly broad: four benchmarks, multiple heterogeneity levels, homogeneous and heterogeneous client models, active-ratio sweeps, noise and label-error robustness tests, and an ablation study. The inclusion of an ablated variant without LPA and the consistency of the main accuracy trends are strengths. However, the theoretical convergence result is not established as a theorem about the actual algorithm: it assumes the key property that LPA is supposed to provide, and it does not connect the per-client local loss bounds to the stated global gradient bound. The experimental claims of 'consistently outperforming' baselines are also weakened by missing hyperparameter values and by several comparisons that are within one standard deviation. The core empirical direction is promising, but the paper currently overclaims both theoretically and empirically.","major_comments":[{"comment":"Assumption 5 postulates precisely the property that LPA is supposed to provide: E[||theta^{k,l}_{t+1} - bar_theta^l_{t+1}||^2] <= delta^2 for k in T_l. The pruning operation in Eq. (12) removes extreme deviations from an average that may itself be corrupted; it does not, by itself, bound the deviation of the remaining elements from that average. Since Eq. (28) inherits delta^2 as an error term, the theorem does not independently establish that LPA improves convergence; at best it states convergence under an unverified premise. Please either prove a bound on the trimmed-mean deviation under an explicit corruption model, or clearly state the result as conditional and provide an independent empirical check of the magnitude of delta^2.","section":"§4.1, Assumption 5 and §4.3, Theorem 1 (Eq. 28)"},{"comment":"The proof jumps from the per-client local-loss decrease bound in Eq. (27) to the global gradient bound in Eq. (28) without a comparison step. The global objective L(theta^g) minimized by the Plug-in update is never defined: Eq. (9) contains only a KL term and no task loss, so it is unclear what stationary point the bound refers to. A valid derivation needs either a virtual global sequence argument or an explicit bound on ||nabla L(theta^g_t)|| in terms of the trusted local gradients. Without one of these, the O(1/sqrt(T)) rate is not a theorem about FedMLAC as specified in Algorithms 1 and 2.","section":"§4.2, Eqs. (27)-(28)"},{"comment":"The paper never reports the values of alpha, v_h, and v_l, which control the mutual-learning trade-off and the LPA pruning fraction. These are central free parameters of the method, and without them the experiments are not reproducible and the comparison across datasets cannot be checked for inadvertent per-dataset tuning. Please report the exact values used for every experiment, state whether they were tuned per dataset, and include a sensitivity analysis for at least alpha and the pruning rates.","section":"§5.2, Hyperparameters and §5.3"},{"comment":"The claim that FedMLAC 'consistently outperforms' all baselines is not supported by the reported error bars on several datasets. In Table 4, FedMLAC on IEMOCAP is 51.97 +/- 5.81 versus 50.48 +/- 5.35 for FedKAD and 48.54 +/- 6.59 for FedAvg; these differences are well within one standard deviation of the runs. The same issue appears in several CREMA-D and Urban Sound rows. Please report paired significance tests or confidence intervals across the five runs, or soften the claim to directional improvement without statistical significance.","section":"§5.3, Table 4"},{"comment":"FedKAD is modified by removing its reliance on public datasets and disabling global-logit distillation, leaving only class-wise feature-map communication. This removes a major component of the original FedKAD method and may disadvantage it, making the reported margins over FedKAD potentially inflated. Please justify this modification with an ablation, or compare against the original FedKAD configuration where feasible, and discuss the effect of the modification on the comparison.","section":"§5.2, Baselines (FedKAD modification)"}],"minor_comments":[{"comment":"The section heading 'Prelimineries' is a typo and should read 'Preliminaries'.","section":"§3.1"},{"comment":"The heterogeneous-model and ablation results are reported only as bar charts, with no numerical table. The text quotes specific improvements (e.g., 'up to 5.23%', 'up to 4.45%'), but readers cannot verify these values or the error bars. Please provide the exact mean +/- standard deviation for every method and configuration.","section":"§5.4 and §5.8, Figures 2 and 3"},{"comment":"Eq. (1) writes the global objective as a sum over L_k(omega_g), but the local loss is later defined as L_k(omega_k); the argument notation is inconsistent and should be aligned.","section":"§3.1, Eq. (1)"},{"comment":"The 'Urban Sound' dataset is referred to informally; the original dataset name 'UrbanSound8K' should be used in the text and citation for clarity and traceability.","section":"§5.1, Datasets"},{"comment":"The sentence 'For other FedAC methods listed in Table 1, fair comparison is not feasible' is vague; please specify which methods were considered and why they could not be included, or remove the sentence if no attempt was made.","section":"§5.2, Baselines"}],"recommendation":"major_revision","confidential_remarks":"The proposed method is closely related to the FedKAD architecture (shared lightweight model plus knowledge distillation), and the current comparison with a modified FedKAD may make the contribution appear incremental. If the convergence section is substantially revised or removed, and the experimental reporting is tightened with statistical tests and full hyperparameter disclosure, the empirical study could be publishable. The current manuscript should not be accepted with the theoretical claim as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it for the empirical story, not the theory. The paper integrates two known ideas—deep mutual learning and layer-wise outlier pruning—into a single federated audio classification framework. Clients train a personalized model and a shared plug-in model with bidirectional KL distillation; the server prunes extreme per-layer deviations before averaging. That specific combination for audio isn't in the prior work I know, and the experiments are broad: four benchmarks, homogeneous and heterogeneous model settings, varying participation, added noise, and label errors. The reported gains are modest (1–3 points) and on IEMOCAP they're inside the standard deviation, so the abstract's \"consistently outperforms\" overstates things. Still, the ablation with LPA disabled shows the pruning component contributes something real, especially under label noise.\n\nThe soft spots are the theory and some practical gaps. Theorem 1 claims an O(1/√T) convergence rate for the global plug-in model, but it doesn't hold as written. Assumption 5 just postulates that LPA keeps trusted-layer deviations bounded by δ², which is basically the result the theorem is supposed to prove. More importantly, the proof jumps from bounds on per-client local plug-in losses to a bound on the gradient of a global objective L(θ^g), but that objective is never defined—Eq. 9 is just a KL loss per client, and aggregation doesn't define an optimization objective for the averaged θ^g. Without a comparison step connecting local gradients to the global gradient at the average, the theorem is not derived from the algorithm. The empirical robustness results are independent of this theorem and stand on their own, so this isn't a load-bearing flaw for the experimental claims, but the theory section shouldn't be cited as proof of convergence.\n\nOther issues: α, v_h, and v_l are never reported, which hurts reproducibility. The FedKAD baseline is modified (public dataset removed), which is defensible but worth flagging. The authors note only the non-convexity and fixed LPA parameters as limitations; the missing objective definition is the bigger gap.\n\nWho's it for: researchers working on personalized or robust federated audio will want this as a baseline. The method is plausible and the experiments are useful. But the theory needs a major rework: either drop the convergence claim or actually define the global plug-in objective and prove a comparison lemma.\n\nMy recommendation: accept for peer review—it's a legitimate integration with real experiments and the issues are fixable—but require the hyperparameters and a rewritten theory section. I'd cite it if I were comparing FL audio methods.","headline":"Solid empirical integration of mutual learning and robust aggregation for federated audio, but the convergence theorem is not established and key hyperparameters are undisclosed.","tokens_in":20834,"tokens_out":3713,"would_cite":true,"duration_ms":39905,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Federated audio model beats baselines under noise, skew, and poison","keywords":["Federated Learning","Audio Classification","Mutual Learning","Knowledge Distillation","Layer-wise Pruning Aggregation","Data Heterogeneity","Model Heterogeneity","Data Poisoning"],"falsifier":"Poison one client's Plug-in model in a single layer while keeping the others clean, then run LPA with the paper's fixed pruning fractions; if the poisoned layer's deviation sits inside the middle subset and the global Plug-in model degrades as much as under plain averaging, Assumption 5 fails in practice. A simpler check is to sweep $v_h$ and $v_l$ from 0 to 30 percent on clean non-IID data: if accuracy peaks sharply only at the hand-picked values and collapses elsewhere, the reported robustness is tuned rather than structural.","tokens_in":1708,"feed_emoji":"🎧","tokens_out":2291,"duration_ms":91539,"temperature":0.7,"pith_summary":"The paper sets out to show that one federated learning framework can handle the three problems that jointly degrade federated audio classification: non-identical data distributions across clients, differing client model architectures, and corrupted or poisoned training labels. It proposes FedMLAC, in which every client trains a personal audio model alongside a small shared Plug-in model, and the two teach each other through bidirectional knowledge distillation. The server then aggregates only the Plug-in updates, using a layer-wise pruning step that discards the most extreme parameter deviations. If the claims hold, FedMLAC gives a single recipe that improves accuracy and robustness on four audio benchmarks while preserving client privacy and architectural flexibility.","feed_headline":"Federated audio model beats baselines under noise, skew, and poison","feed_subtitle":"Mutual learning plus per-layer pruning lets heterogeneous clients share knowledge without sharing raw audio.","key_machinery":"The load-bearing object is the lightweight, globally shared Plug-in model, which decouples personalization from aggregation. It is trained on each client together with the client's private model using mutual learning: the local model's task loss is augmented with a KL-divergence term that aligns it with the Plug-in model's predictions, and the Plug-in model in turn is updated with a KL loss that aligns it with the local model, so knowledge flows both ways. The second mechanism is Layer-wise Pruning Aggregation (LPA), which sorts each layer's $\\ell^2$ deviation from the cohort average and prunes the largest $\\lfloor v_h |\\mathcal{S}_t|\\rfloor$ and smallest $\\lfloor v_l |\\mathcal{S}_t|\\rfloor$ clients before weighted averaging. LPA's effectiveness enters the convergence proof as a bounded-deviation assumption, and the theorem states that the global Plug-in model reaches a stationary point at rate $\\mathcal{O}(1/\\sqrt{T})$.","core_discovery":"FedMLAC claims that data heterogeneity, model heterogeneity, and data poisoning can be addressed simultaneously rather than piecemeal. The mechanism is a decoupled design: each client keeps a personalized local audio classification model for its own data and a lightweight global Plug-in model; bidirectional KL-divergence distillation transfers global knowledge to the client model and client knowledge to the Plug-in model. On the server, Layer-wise Pruning Aggregation (LPA) computes each layer's $\\ell^2$ deviation from the cohort average, removes the top and bottom fraction of deviating updates, and averages the trusted remainder, so outlier updates are filtered before they enter the global model. The paper reports consistent accuracy gains over FedAvg, FedProx, FedOPT, and FedKAD in both homogeneous and heterogeneous model settings, smaller degradation under additive noise and label errors, and an $\\mathcal{O}(1/\\sqrt{T})$ convergence bound for the global Plug-in model.","pith_inferences":["Editorial inference: if per-layer trimmed averaging works as claimed, the same operation could harden other federated aggregation schemes in vision or language, wherever server-side averaging is the bottleneck.","Editorial inference: the paper fixes the pruning fractions $v_h$ and $v_l$; an adaptive per-layer or per-round choice would be a testable extension, since the optimal trusted fraction likely depends on how many clients are actually poisoned.","Editorial inference: because only the Plug-in model is exchanged, the framework may cut communication cost relative to full-model averaging, but the paper does not quantify bytes per round, so that saving is a measurable follow-up rather than an established result."],"forward_implications":["FedMLAC is reported to outperform FedAvg, FedProx, FedOPT, and FedKAD on GSC, IEMOCAP, CREMA-D, and Urban Sound, in both homogeneous and heterogeneous model settings.","Under white Gaussian noise down to 10 dB SNR and label error rates up to 0.5, FedMLAC degrades less than every baseline and less than its own LPA-free variant, indicating that LPA carries the robustness gain.","Because clients upload only the lightweight Plug-in model, the framework supports heterogeneous client architectures and partial participation without needing a shared public dataset for distillation.","Theorem 1 states that with learning rate proportional to $1/\\sqrt{T}$, the global Plug-in model's average squared gradient norm converges to a stationary point at rate $\\mathcal{O}(1/\\sqrt{T})$, with the LPA deviation bound $\\delta^2$ appearing only in the error term."],"supporting_citations":[{"why":"supplies the FedAvg aggregation rule in Equation (3) and serves as the primary homogeneous-model baseline.","marker":"[35]"},{"why":"provides the deep mutual learning objective with bidirectional KL divergence that FedMLAC-Update builds on.","marker":"[36]"},{"why":"is the FedKAD baseline, the knowledge-distillation method designed for data and model heterogeneity that FedMLAC is compared against.","marker":"[34]"},{"why":"supplies the FedAudio benchmark protocol, preprocessing choices, and the label-noise injection setup used across experiments.","marker":"[41]"},{"why":"is the FedProx baseline for non-IID data via a proximal regularization term.","marker":"[43]"},{"why":"is the FedOPT adaptive federated optimization baseline.","marker":"[44]"},{"why":"provides the Google Speech Commands dataset for keyword spotting experiments.","marker":"[37]"},{"why":"provides the Urban Sound dataset for environmental sound event classification.","marker":"[38]"},{"why":"provides the IEMOCAP dataset for emotional speech recognition experiments.","marker":"[39]"},{"why":"provides the CREMA-D dataset for emotion recognition experiments.","marker":"[40]"}],"fun_headline_variants":["One federated audio model: three challenges solved at once","Mutual learning plus layer pruning: robust federated audio","Federated audio: pruning filters poison, distillation beats skew","Heterogeneous clients share audio knowledge without sharing data","FedMLAC: one framework for skew, poison, and model diversity"],"cache_read_input_tokens":23040,"weakest_assumption_plain":"The convergence proof rests on Assumption 5, which says that after Layer-wise Pruning Aggregation prunes the extreme layers, the trusted clients' parameters stay within a bounded deviation $\\delta^2$ of the cohort average; if pruning does not actually keep deviations bounded in practice, the theorem's error term and part of the robustness story do not follow.","fun_headline_variants_meta":{"raw":{"variants":["One federated audio model: three challenges solved at once","Mutual learning plus layer pruning: robust federated audio","Federated audio: pruning filters poison, distillation beats skew","Heterogeneous clients share audio knowledge without sharing data","FedMLAC: one framework for skew, poison, and model diversity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000413,"raw_usage":{"total_tokens":2122,"prompt_tokens":919,"completion_tokens":1203,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":1119}},"tokens_in":535,"tokens_out":1203,"duration_ms":9892,"temperature":1.0,"reasoning_tokens":1119,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:32:03.216713+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Poison one client's Plug-in model in a single layer while keeping the others clean, then run LPA with the paper's fixed pruning fractions; if the poisoned layer's deviation sits inside the middle subset and the global Plug-in model degrades as much as under plain averaging, Assumption 5 fails in practice. A simpler check is to sweep $v_h$ and $v_l$ from 0 to 30 percent on clean non-IID data: if accuracy peaks sharply only at the hand-picked values and collapses elsewhere, the reported robustness is tuned rather than structural.","supporting_citations":[{"cited_title":"McMahan, E","cited_arxiv_id":null,"evidence_quote":"supplies the FedAvg aggregation rule in Equation (3) and serves as the primary homogeneous-model baseline."},{"cited_title":"Zhang, T","cited_arxiv_id":null,"evidence_quote":"provides the deep mutual learning objective with bidirectional KL divergence that FedMLAC-Update builds on."},{"cited_title":"Zhang, T","cited_arxiv_id":null,"evidence_quote":"supplies the FedAudio benchmark protocol, preprocessing choices, and the label-noise injection setup used across experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the FedProx baseline for non-IID data via a proximal regularization term."},{"cited_title":"Salamon, C","cited_arxiv_id":null,"evidence_quote":"provides the Urban Sound dataset for environmental sound event classification."},{"cited_title":"Busso, M","cited_arxiv_id":null,"evidence_quote":"provides the IEMOCAP dataset for emotional speech recognition experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the CREMA-D dataset for emotion recognition experiments."}],"review_version":1}