{"id":"281b540b-2e0f-4766-a845-537f407dce44","arxiv_id":"2412.11674","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UA-PDFL measures client data skew using model outputs on a fixed unit input and switches between layer-wise personalization and whole-model replacement, improving decentralized federated learning accuracy on non-IID data.","lead":"This paper proposes UA-PDFL, a way to train models across many devices without a central server, where each device keeps part of its model private. It uses a fixed 'unit' input to measure data differences between devices and then shares only the right layers, improving accuracy on uneven data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unit-representation divergence in Eqs. (5)–(7) is the load-bearing proxy for client data similarity, yet Eq. (3) admits the proxy is unequal to the true distribution; if Xunit outputs encode model artifacts, both dropout and personalization decisions are noise-driven.","rationale":"The reader's weakest assumption is exactly the unit-representation premise, and I agree. The paper's own Eq. (3) is the strongest evidence that the premise is unestablished: it concedes the inference distribution on an unknown sample is not the data distribution, and the claim that Xunit ∈ Xu fixes this is a non-sequitur—membership in the sample space does not make a fixed all-ones tensor representative of any client's input distribution. The empirical sections provide breadth (four datasets, three architectures, two skew levels) and the reported gains are plausible, but they do not isolate the unit representation as the cause; the §5.5 ablation compares only extreme cases. Because the threshold thI and coefficient µ are not reported, the experiments cannot be reproduced exactly, compounding the difficulty of testing the premise post hoc. The proposed check is inexpensive and decisive: if Div on Xunit does not rank-order known distributional similarities, the central mechanism is not sound, and the paper would need a different distribution proxy, or a direct validation of Xunit, before acceptance. Since the reader already conditioned acceptance on resolving issues of this kind, I do not move the verdict; I would make validation of Div on partial-overlap and equal-marginal pairs an explicit acceptance criterion.","tokens_in":20841,"tokens_out":6110,"duration_ms":63317,"concrete_test":"Construct a 30-client CIFAR-10 Dirichlet partition and add a controlled pair set: two clients with identical label marginals but different input domains (e.g., one applies random horizontal flips and color jitter to its local images), two with partially overlapping labels, and two with disjoint labels. At round 50, compute Div(i,j) via Xunit and compare, over all pairs, to a reference divergence (e.g., true JS divergence between empirical label distributions or output disagreement on a small held-out public set). If the Spearman rank correlation is below about 0.8, or if the identical-marginal/different-domain pair is not ranked among the most similar pairs, the unit-representation premise fails. Rerunning UA-PDFL with Xunit replaced by random noise would then show whether the reported gains are an artifact of the chosen fixed input.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that UA-PDFL adapts personalization to varying data skew—depends on Div(i,j) computed from model outputs on the fixed all-ones tensor Xunit (Eqs. 5–7) being a faithful measure of the similarity of clients' real data distributions. This premise is asserted ('as Xunit ∈ Xu, it can inherently and accurately represent the data distribution differences across clients') but not proven. The paper itself states in Eq. (3) that p(Dm) ≠ p(σ(fm(wm; Xu))), so the output on a single out-of-distribution tensor cannot be assumed to represent the joint input-label distribution; at best it is a function of the model's predictions for one point. A concrete failure mode: two clients can have identical label marginals but different conditional input distributions (e.g., same classes from different domains); their Xunit outputs could be nearly identical, so UA-PDFL would treat them as similar and aggregate classifiers that should stay separate. Conversely, output divergence can be driven by architecture, initialization, or out-of-distribution behavior rather than by data. The supporting evidence is limited to a three-client toy experiment with disjoint label sets (§3.2) and an ablation (§5.5) that compares only identical versus distinct distributions; there is no sensitivity analysis over Xunit values, no comparison to a ground-truth distributional divergence, and no test on partial-overlap clients. Since Algorithm 2 gates every aggregation and communication decision on Div(i,j) (lines 10–16), any failure of this proxy invalidates the claimed adaptive mechanism, not just a tuning detail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UA-PDFL, a decentralized federated learning method in which each client computes a 'unit representation' by feeding a fixed all-ones tensor through its local model, uses Jensen-Shannon divergence between these representations to decide client-wise dropout and layer-wise personalization, and adds an auxiliary feature-representation regularization term to the local objective. The method is evaluated on CIFAR-10, CIFAR-100, SVHN, and PathMNIST with three architectures and two Dirichlet heterogeneity levels, and is compared against five baselines plus an ablation. The paper claims to be the first to adaptively adjust the number of personalization layers via unit representations without relying on public data.","tokens_in":21199,"tokens_out":4882,"duration_ms":47566,"significance":"If the empirical results hold, UA-PDFL would offer a serverless FL approach that adapts personalization structure to the degree of data skew, avoids public data, and reduces communication overhead under relatively IID conditions. The experimental breadth is a genuine strength: four datasets, three architectures, two heterogeneity regimes, an extreme non-IID case, communication-cost measurements, and an ablation study. However, the load-bearing unit-representation proxy is not validated beyond a toy example and an incomplete ablation, and the convergence analysis in Section 4 does not model the actual algorithm. The empirical claims are plausible but not yet backed by a rigorous mechanism analysis, so the paper needs substantial revision before the central claims are fully supported.","major_comments":[{"comment":"The central premise that Div(i,j) computed from unit representations on the all-ones tensor Xunit is a faithful proxy for the similarity of client data distributions is not established. Eq. (3) explicitly concedes that p(Dm) is not equal to p(σ(fm(wm; Xu))), and the subsequent statement that 'as Xunit ∈ Xu, it can inherently and accurately represent the data distribution differences across clients' is asserted without proof. Div(i,j) is a function of the model's output vector for a single out-of-distribution input; it can be insensitive to changes in conditional input distributions when label marginals coincide, and it can be driven by initialization, architecture, or out-of-distribution behavior. Since Algorithm 2 gates both dropout and layer-wise personalization on this metric, a failure of this premise makes the mechanism noise-driven. The toy experiment in Fig. 4 and the ablation in Fig. 12 only compare disjoint-label versus identical distributions. I would require a direct validation against a ground-truth distributional divergence, a sensitivity analysis over the entries aij, and at least one partial-overlap or same-marginal/different-domain scenario.","section":"Section 3.2, Eqs. (5)-(7)"},{"comment":"The convergence analysis does not analyze the algorithm presented in Algorithms 1-2. Eq. (19) asserts Var(ξr) ≤ σ²/M 'by applying client dropout' with no derivation, and no connection is made to the threshold thI or to the dropout rule in Algorithm 2, which replaces the model with a random client only when all divergences are below thI and is inactive under heterogeneous data, precisely the setting where the empirical gains are claimed. The auxiliary representation term is dropped by asserting ∥I_Aux_m − I_Aux_avg∥² → 0 without proof, even though Eq. (12) makes it part of the local objective. Furthermore, the derivation tracks a single global parameter w, whereas the method maintains per-client feature extractors and classifiers that are aggregated differently. As written, Eq. (28) is a standard SGD convergence bound for a different algorithm; please either supply a proof for the actual UA-PDFL updates or clearly label Section 4 as a simplified informal analysis that does not make claims about the full method.","section":"Section 4, Eq. (19) and Eq. (28)"},{"comment":"The ablation does not demonstrate that the unit representation, rather than the mechanism structure, is responsible for the reported gains. Full UA-PDFL is compared only against removing client-wise dropout or removing layer-wise personalization; there is no control that replaces Div(i,j) with random client selection or with a different similarity metric, and no sensitivity analysis for the threshold thI or the coefficient µ. Since thI determines when dropout replaces the local model and when classifiers are aggregated, the central claim that the method 'adaptively adjusts' personalization to data skew is not supported without showing how performance varies with thI and with the unit tensor entries aij. This is fixable by adding a small sensitivity study and a random-proxy baseline.","section":"Section 5.5 and Algorithm 2"}],"minor_comments":[{"comment":"The computation of I_Aux_avg is inconsistent: Algorithm 1 line 17 appears to use only clients satisfying Div(m,i) < thI, while Algorithm 2 line 28 uses every client in Qi regardless of divergence; please harmonize the pseudocode and the surrounding text.","section":"Algorithms 1 and 2"},{"comment":"The tables report mean±std but the number of independent runs is not stated; please provide the seed count and evaluation protocol so the reader can assess the stability of the comparisons.","section":"Section 5.1.4"},{"comment":"The communication-cost figure reports values in units of 1e11 but the text does not explain how parameter transmission volume is counted (e.g., number of parameters × rounds × clients); please clarify the metric.","section":"Figure 11"},{"comment":"There are numerous typos and English errors, such as 'an Unit representation' in Section 3, 'mantains' in Section 3.4, 'the the local loss' near Eq. (12), 'governs governs' in Section 5.1.2, and 'Leaning momentum' in Section 5.1.4; these should be corrected.","section":"Throughout"},{"comment":"The claim that UA-PDFL is 'superior or comparable' is based on final accuracy differences that are often within one standard deviation; the authors should avoid overstating small differences and ideally report pairwise significance tests or effect sizes.","section":"Section 5.2, Tables 1-2"}],"recommendation":"major_revision","confidential_remarks":"The novelty framing as 'the first work' to introduce unit representation should be checked against existing inference-similarity and proxy-based methods in the FL literature; if related work exists, the contribution statement needs to be more precise. I would also encourage the editor to request code release, since the paper's empirical claims are extensive and a public implementation would greatly aid reproducibility and allow the community to stress-test the unit-representation premise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: UA-PDFL is an empirical method paper with a genuinely new ingredient — using a fixed all-ones tensor as a shared probe to estimate client data skew without a server or public data — and its evaluation is broader and more honest than most DFL papers. But the central mechanism is supported by assertion more than evidence, and the convergence section doesn't analyze the actual algorithm. It deserves a serious referee, with revision required.\n\nWhat's new: the specific combination of unit representation, client-wise dropout, and layer-wise personalization in a decentralized protocol is not in the cited prior work. The unit representation is a close relative of FLIS's inference-similarity clustering, but the fixed unit tensor is a clever simplification that removes the need to negotiate a reference input. The experiments cover four datasets, three architectures, two heterogeneity levels, and a real-world medical image set; the gains over strong baselines are usually modest (1-3 points) but consistent, and the communication-cost analysis is a genuine plus. The ablation shows both mechanisms contribute, and the authors acknowledge the higher communication cost of DFL in their conclusion.\n\nSoft spots, in order of severity. First, the load-bearing assumption: Div(i,j) computed from model outputs on Xunit is treated as a faithful measure of the similarity of clients' true data distributions. The paper itself concedes in Eq. (3) that the output distribution is not the real distribution. The evidence offered is a three-client toy example with disjoint label sets and an ablation that only compares identical versus distinct distributions. There is no sensitivity analysis over Xunit entries, no test with partial label overlap, and no comparison against a ground-truth divergence. Since Algorithm 2 gates every aggregation and communication decision on this divergence, a failure here is not a tuning detail. This is the main thing I would push on.\n\nSecond, the convergence analysis is detached from the algorithm. Eq. (19) asserts Var(xi_r) <= sigma^2/M 'by applying client dropout' with no derivation, and the auxiliary representation term is dropped by fiat. The resulting bound is a generic FL convergence bound, not a statement about UA-PDFL's switching dynamics. It should either be derived properly or cut.\n\nThird, key hyperparameters thI and mu are never reported, so exact reproduction is impossible. Fourth, the abstract's claim about 'adaptively adjusting the level of personalization layers' overstates what the algorithm does — it switches between two modes and never changes the split point between feature extractor and classifier.\n\nWho this is for: DFL and edge/IoT researchers looking for a practical personalization heuristic will get value. Theory readers won't. I'd send it to review, with the demand that the unit-representation premise get a sensitivity analysis and a partial-overlap test, the convergence theorem get fixed or removed, and the missing hyperparameters be reported.","headline":"A practical DFL personalization method with a genuinely new unit-representation trick and broad experiments, but the load-bearing proxy is under-validated and the convergence proof does not analyze the actual algorithm.","tokens_in":21722,"tokens_out":2927,"would_cite":true,"duration_ms":28870,"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":"UA-PDFL claims adaptive personalization guided by a fixed unit tensor makes decentralized federated learning robust to data skew without public data or a central server.","keywords":["decentralized federated learning","personalized federated learning","unit representation","non-IID data","client-wise dropout","layer-wise personalization","data heterogeneity"],"falsifier":"Train two clients on the same data distribution with different random seeds or different architectures, then compute their unit representations and $\\mathrm{Div}(i,j)$; if any such pair exceeds the paper's threshold $th_I$, the unit representation is not a faithful probe of true distribution similarity.","tokens_in":20624,"feed_emoji":"🧠","tokens_out":6112,"duration_ms":58462,"temperature":0.7,"pith_summary":"This paper tries to make decentralized federated learning work when client data are unevenly distributed, without a central server and without any public data. It proposes UA-PDFL, in which every client feeds the same fixed tensor, all entries equal to one, through its local model and treats the output probability vector as a \"unit representation\" of that client's data distribution. The Jensen-Shannon divergence between these vectors decides which clients are similar enough to aggregate, and the number of shared feature-extractor layers is adapted per client rather than fixed. If the claim holds, a serverless federation can tune its personalization strength to the actual degree of data skew and stay accurate where fixed-personalization baselines falter.","feed_headline":"One all-ones tensor tells clients how to personalize","feed_subtitle":"UA-PDFL adapts sharing to data skew, matching or beating five baselines on four image datasets without a central server.","key_machinery":"The load-bearing object is the unit representation $I_m = \\sigma(f_m(w_m; X_{\\mathrm{unit}}))$: the softmax output each client's model produces on the fixed unit tensor $X_{\\mathrm{unit}}$, all entries equal to one. Pairwise Jensen-Shannon divergence $\\mathrm{Div}(i,j)$ computed from these vectors drives every design choice: whether to drop clients when all divergences fall below threshold $th_I$, which classifiers to fuse, and which feature extractors to aggregate. A second auxiliary representation $I^{\\mathrm{Aux}}_m = g_m(X_{\\mathrm{unit}})$ from the feature extractor feeds a proximal regularizer $\\mu\\,\\|I^{\\mathrm{Aux}}_m - I^{\\mathrm{Aux}}_{\\mathrm{avg}}\\|_2^2$ that pulls local extractors toward a shared feature space while classifiers stay personalized. The paper also justifies client-wise dropout as a Bayesian approximation, viewing each connected client as a neuron and the aggregation step as a linear combination.","core_discovery":"The paper's central claim is that a single hand-set input, the unit tensor, can act as a distribution probe: feeding $X_{\\mathrm{unit}}$ with all entries equal to $1$ to each client model and comparing the resulting softmax outputs with Jensen-Shannon divergence yields a metric $\\mathrm{Div}(i,j)$ that ranks client-pair similarity in line with their true data distributions. On top of this metric the paper builds two mechanisms: client-wise dropout, which replaces a client's model with a randomly chosen connected client's model when all divergences are below a threshold, and layer-wise personalization, which aggregates feature extractors among similar clients while keeping each client's final model a weighted combination of similar peers' classifiers. A proximal term on auxiliary representations from the feature extractor steers local extractors toward a common feature space. The experiments on CIFAR-10, CIFAR-100, SVHN, and PathMNIST show UA-PDFL matching or beating five baselines and converging faster in most settings, with the largest gains on the hardest skewed task, CIFAR-100.","pith_inferences":["Because $\\mathrm{Div}$ lives in output-probability space, the method implicitly assumes all clients share the same label set and output layer; clients with different architectures or label spaces would need an agreed projection before the divergence is meaningful.","The unit representation is a single fixed probe, so its reliability may depend on model calibration; a natural test is whether averaging over many random fixed probes or using feature-map statistics makes the divergence more stable across architectures.","Client-wise dropout in the near-IID regime effectively turns training into mini-batch SGD on one random client per round, so for homogeneous data the method's benefit may reduce to communication savings rather than accuracy gains.","A straightforward extension would replace the hand-set threshold $th_I$ with an adaptive or per-client threshold derived from the divergence distribution, removing the method's most sensitive hyperparameter."],"forward_implications":["If UA-PDFL is right, decentralized federated learning no longer needs public data or a server to detect client similarity: a fixed dummy input and the Jensen-Shannon divergence between output vectors suffice.","Adaptive personalization depth should outperform fixed personalization layers on both ends of the skew spectrum, since client-wise dropout handles near-IID clients while layer-wise personalization handles strong non-IID.","The largest accuracy gains appear on the hardest heterogeneous tasks, CIFAR-100 and extreme PathMNIST non-IID, suggesting the mechanism is most valuable when client distributions overlap least.","Communication volume is not consistently reduced: the paper's own measurements show UA-PDFL communicates more than DFedAvgM and DisPFL, with savings from client-wise dropout appearing mainly under more uniform data."],"supporting_citations":[{"why":"Supplies the federated learning formulation, the weighted-averaging baseline, and the extreme non-IID data partitioning method used in experiments.","marker":"[32]"},{"why":"Supplies the base-layer versus personalization-layer decomposition that UA-PDFL generalizes into adaptive layer-wise personalization.","marker":"[1]"},{"why":"Supplies the similarity-guided decentralized connection idea and the convergence-analysis scaffolding the paper builds on.","marker":"[30]"},{"why":"Serves as the main decentralized personalized sparse-training baseline that UA-PDFL compares against.","marker":"[5]"},{"why":"Supports the premise that parametric models encode training-data information, which motivates replacing public data with a unit tensor.","marker":"[57]"},{"why":"Defines the Jensen-Shannon divergence used as the pairwise client-similarity metric on unit representations.","marker":"[33]"},{"why":"Provides the dropout-as-Bayesian-approximation argument that justifies client-wise dropout in the decentralized setting.","marker":"[8]"},{"why":"Serves as the momentum-based decentralized baseline whose communication efficiency is measured against UA-PDFL.","marker":"[45]"}],"fun_headline_variants":["A single all-ones tensor gauges data skew for personalized federated learning","Unit tensor probe adapts personalization in decentralized federated learning","UA-PDFL uses an all-ones tensor to tune client personalization","One tensor to measure skew: personalized DFL without a server","All-ones tensor reveals client similarity for adaptive personalization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire method rests on the assumption that a fixed all-ones input, pushed through each client's model, produces a unit representation whose pairwise Jensen-Shannon divergence reliably tracks how different the clients' real data distributions are.","fun_headline_variants_meta":{"raw":{"variants":["A single all-ones tensor gauges data skew for personalized federated learning","Unit tensor probe adapts personalization in decentralized federated learning","UA-PDFL uses an all-ones tensor to tune client personalization","One tensor to measure skew: personalized DFL without a server","All-ones tensor reveals client similarity for adaptive personalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3370,"prompt_tokens":1002,"completion_tokens":2368,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2277}},"tokens_in":618,"tokens_out":2368,"duration_ms":13902,"temperature":1.0,"reasoning_tokens":2277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:41:54.149621+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two clients on the same data distribution with different random seeds or different architectures, then compute their unit representations and $\\mathrm{Div}(i,j)$; if any such pair exceeds the paper's threshold $th_I$, the unit representation is not a faithful probe of true distribution similarity.","supporting_citations":[{"cited_title":"Like attracts like: Personalized federated learning in decentralized edge computing","cited_arxiv_id":null,"evidence_quote":"Supplies the federated learning formulation, the weighted-averaging baseline, and the extreme non-IID data partitioning method used in experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the base-layer versus personalization-layer decomposition that UA-PDFL generalizes into adaptive layer-wise personalization."},{"cited_title":"Fedcoin: A peer-to-peer payment system for federated learning, in: Federated learning: privacy and incentive","cited_arxiv_id":null,"evidence_quote":"Supplies the similarity-guided decentralized connection idea and the convergence-analysis scaffolding the paper builds on."},{"cited_title":"Dispfl: Towards communication-efficient personalized federated learning via decentral- ized sparse training, in: International conference on machine learning, PMLR","cited_arxiv_id":null,"evidence_quote":"Serves as the main decentralized personalized sparse-training baseline that UA-PDFL compares against."},{"cited_title":"The jensen- shannon divergence","cited_arxiv_id":null,"evidence_quote":"Defines the Jensen-Shannon divergence used as the pairwise client-similarity metric on unit representations."},{"cited_title":"Dropout as a bayesian approximation: Insights and applications, in: deep learning workshop, ICML, p","cited_arxiv_id":null,"evidence_quote":"Provides the dropout-as-Bayesian-approximation argument that justifies client-wise dropout in the decentralized setting."},{"cited_title":"Decentralized federated averaging","cited_arxiv_id":null,"evidence_quote":"Serves as the momentum-based decentralized baseline whose communication efficiency is measured against UA-PDFL."}],"review_version":1}