{"id":"cd94fffa-a1b2-4841-9d37-e71dfa66d659","arxiv_id":"2505.02426","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A literature survey of one-shot federated learning that organizes methods, datasets, and code, but reports no new experimental or theoretical results.","lead":"This paper surveys one-shot federated learning, a way to train a shared AI model across many devices using a single round of communication instead of many. It categorizes current methods, lists open-source tools, and discusses open problems such as data diversity, privacy, and scalability.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Survey's comprehensiveness claim is undercut by irreconcilable PRISMA counts and citation/category errors, but the taxonomy survives as a corrigible entry-point.","rationale":"The reader's weakest_assumption was completeness and accuracy of the literature search and taxonomy; I agree and found additional internal evidence (the unreconciled 94-to-79 count, Tiny-ImageNet misattribution, Andrew et al. under adversarial robustness, and the missing Section 4). These are correctness risks for a survey's central promise, but they are fixable and do not invalidate the taxonomy or the future-directions discussion. Hence CONDITIONAL is the correct verdict: acceptable after a careful revision, not as a final 'comprehensive reference'. I do not recommend REJECT because the paper does provide a usable structured overview, a curated code table, and a reasonable challenge list; the flaws undermine completeness but not the core survey utility.","tokens_in":22614,"tokens_out":2413,"duration_ms":26076,"concrete_test":"Reconstruct the PRISMA flow: re-run the exact query on Scopus, arXiv and OpenReview, deduplicate titles, and count how many of the 79 selected papers appear in the union. Then check whether papers discussed in Section 3.2 that are absent from the PRISMA set exceed a small threshold (e.g., 5). Separately, verify the Tiny-ImageNet citation and confirm whether Andrew et al. (2023) is tested under an adversarial-robustness benchmark or is purely a privacy-estimation study; if it is the latter, move it to a privacy category and recheck the Section 3.2.4 taxonomy. This settles whether the survey's map is complete and correctly labeled or merely an oversampled convenience set.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's load-bearing assertion is that it is a 'comprehensive reference' that 'systematically' maps One-shot FL. That claim stands or falls on the completeness of the literature collection (Section 2.1) and on the accuracy of the taxonomy entries in Section 3.2. Internal evidence shows both are imperfect. First, the screening flow is arithmetically inconsistent: Section 2.1.3 reports 73 papers from databases plus 21 from manual search = 94 initial; after removing duplicates and review articles it says 79 retained; then it says '79 papers were finally selected' with no full-text exclusion step documented. If all 79 survived screening, the flow has no attrition, which is suspicious for a PRISMA-style review; if 15 were removed, the reported number is wrong. Second, Table 2 attributes Tiny-ImageNet to Krizhevsky et al. (2009), a CIFAR paper, and the dataset-description paragraph repeats the same attribution. Third, Section 3.2.4 ('Adversarial Robustness') includes Andrew et al. (2023), which estimates differential-privacy loss under the Gaussian mechanism; that is a privacy-estimation paper, and the paper does not explain how it is an adversarial-robustness method. Fourth, the promised Section 4 (open-source tools) never appears; the text jumps from 3.4 to 5, despite Section 1.5 and Figure 7 referencing it. Each error is small, but together they mean the reader cannot fully trust the map: the count of the field cannot be reproduced from the reported flow, a dataset reference is wrong, a taxonomy bucket is missorted, and a promised resource section is missing. The useful content (taxonomy, challenges, benchmark proposal) is still there, but the 'comprehensive reference' claim is overstated until these are fixed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of one-shot federated learning (One-shot FL). It motivates the single-round communication setting as a response to communication and privacy costs in classical FL, presents a mathematical formulation of the server-side aggregation problem, and organizes recent work into a taxonomy of four categories: theory-based optimization algorithms, knowledge-distillation-based ensemble methods, data heterogeneity, and adversarial robustness. It also discusses applications in Web3.0, autonomous vehicles, IoT, healthcare, recommender systems, and LEO satellite constellations, lists datasets and open-source code repositories, and closes with challenges and future directions including a proposed benchmark. The paper contains no new experimental or theoretical results; its contribution is intended to be a systematic, comprehensive map of the field.","tokens_in":22944,"tokens_out":6186,"duration_ms":70085,"significance":"If accurate, the survey would be a useful entry point for researchers entering one-shot FL: it collects many recent papers, provides a compact notation for the ensemble-distillation framework, and offers practical pointers to datasets and repositories. The strengths are the breadth of covered literature and the explicit resource lists. However, the paper's value is inseparable from the reliability of its map: readers rely on correct category assignments, correct dataset references, and a reproducible literature-collection procedure. The inconsistencies identified below, an irreproducible PRISMA-style count, a dataset misattribution, a misplaced privacy paper in the adversarial-robustness category, and inconsistent section numbering, undercut the comprehensiveness claim as currently stated. These are fixable within the manuscript's scope.","major_comments":[{"comment":"The reported screening flow is internally inconsistent and cannot be reproduced. The text states that 73 papers were identified by database search and 21 by manual search, giving 94 initial papers; after removing duplicates and review articles, 79 were retained; it then states that '79 papers were finally selected.' No full-text eligibility step or exclusion count is reported, so either all 79 survived full-text screening, which would be unusual for a PRISMA-style flow, or one of the counts is wrong. Since the abstract's 'thorough examination' and Section 2's claim to follow PRISMA are the paper's central evidentiary basis, the authors must provide a complete PRISMA-style flow diagram with exclusion reasons and reconcile the arithmetic. Without this, the claimed comprehensiveness cannot be checked.","section":"Section 2.1.3"},{"comment":"Tiny-ImageNet is attributed to Krizhevsky et al. (2009) in both the text of Section 5.1 and Table 2. Krizhevsky et al. (2009) is the CIFAR technical report and does not describe Tiny-ImageNet; Tiny-ImageNet is a derivative of ImageNet with a separate provenance. This is a factual misattribution in the dataset catalog, which is one of the paper's claimed contributions. Please replace it with the correct reference or mark the dataset as introduced by the Tiny ImageNet project, and check the surrounding dataset rows for similar errors.","section":"Section 5.1 and Table 2"},{"comment":"The section titled 'Adversarial Robustness' includes Andrew et al. (2023) as one of its three entries. Andrew et al. (2023) proposes one-shot empirical privacy estimation for differentially private federated learning under the Gaussian mechanism; it is a privacy-estimation paper, not a method for adversarial robustness or backdoor defenses. The same paper is discussed appropriately in Section 3.3 in a privacy context. This mismatch suggests the taxonomy categories are not carefully enforced and makes the classification of methods unreliable for readers. Either move the entry to a privacy-related category or add an explicit justification of its relevance to adversarial robustness.","section":"Section 3.2.4"},{"comment":"The paper's internal structure does not match its own roadmap. Section 1.5 promises that Section 4 will cover open-source code and tools and that Section 5 will cover challenges; the actual text jumps from Section 3.4 to Section 5, where the content is datasets and open-source codes, and Section 6 contains the challenges. Figure 7's caption refers to 'Section c.' This mismatch is more than a typo: it prevents readers from locating the promised discussion and suggests the manuscript was assembled from parts with inconsistent numbering. The section labels, the roadmap, and the figure cross-references need to be unified.","section":"Sections 1.5, 3.4, 5, and Figure 7"}],"minor_comments":[{"comment":"The prose in Section 3.1 needs editing: 'Consider there have a set of clientsC' is ungrammatical, the text after Eq. (2) beginning 'Aw(x); {θk}nk=1 means...' is incomplete, and 'the θk-th client' should be 'the k-th client.'","section":"Section 3.1"},{"comment":"Several typos should be fixed: 'Knowledge Disllation' in the Section 3.2.2 heading, 'Comparsion' in the Table 1 caption, 'NO-IID' in a bullet in Section 3.2.4, and 'NoIID-Bench' in Section 6.1.1, which should be 'NIID-Bench.'","section":"Sections 3.2.2, Table 1, and Section 6.1.1"},{"comment":"In Section 6.1.4, 'Mendieta et al. demonstrated that...' lacks a year, and the corresponding reference in the bibliography is incomplete: 'Mendieta, M., Sun, G., Chen, C., . Exploring...' Provide the full citation.","section":"Section 6.1.4"},{"comment":"The Table 3 entry 'IntactOne Shot FL' has a repository URL containing a space and will not resolve; also, the table mixes papers discussed in the text with others such as Dennis et al. (2021) and Garin and Quintana (2023) that are not described in Section 3, so the correspondence between the taxonomy and the code list is unclear.","section":"Table 3 and Section 5.4"},{"comment":"The text says 'This paper presents a feasible and comprehensive benchmarking framework for One-shot FL,' but no concrete benchmark implementation or evaluation is provided; consider softening this to 'we outline requirements for a benchmarking framework.'","section":"Section 6.3.4"},{"comment":"Section 2.2 says one-shot FL was introduced at ICML 2020, but the cited foundational reference is Guha et al. (2019); clarify whether the introduction is the arXiv preprint or the ICML version to avoid confusion.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading this one. First, it is a real survey with a usable taxonomy, curated code/dataset lists, and a sensible benchmark proposal; it is not a research claim, and it does not pretend to be. Second, the paper's 'comprehensive' claim is undercut by its own reporting: the PRISMA-style flow is arithmetically odd, a promised section is missing, and a few citation/category errors appear exactly where a reader would rely on accuracy. None of that kills the paper, but it needs a careful revision. What is actually new is mostly organizational. The taxonomy into theory-based optimization, knowledge-distillation ensembles, data heterogeneity, and adversarial robustness is reasonable, and it matches how the field already talks about itself. The open-source code table and dataset list are the most concrete value: they give a newcomer places to start. The equations in Section 3.1 are standard data-free distillation formalism, correctly restated. The benchmark recommendation in Section 6.3.4 is a decent sketch, though it is a proposal, not an implemented benchmark. The soft spots are real but proportionate. The screening flow in Section 2.1.3 says 73 database hits plus 21 manual hits equals 94, then 79 retained after dedup/review removal, then 79 are 'finally selected'—meaning no attrition from screening to inclusion. Either the writing is sloppy or the flow is fabricated; a PRISMA-style survey should show that step. Section 4, promised in Section 1.5 and referenced by Figure 7, does not exist; the open-source content lives in Section 5. Table 2 attributes Tiny-ImageNet to Krizhevsky et al., which is the CIFAR reference, and the text repeats the error. Section 3.2.4 groups Andrew et al. (2023), a privacy-loss estimation paper, under adversarial robustness; it fits better under trustworthiness or privacy. These are all fixable, and fixing them matters because they appear in the parts of the paper that a user would actually consult. The central value—an organizing map with resources—holds up. If you are not in one-shot FL, this will not change your life, but if you need to know who did what and where the code is, it is a convenient starting point. I would not cite it as an authoritative reference until the errors are corrected, but I would bring it to a reading group interested in communication-efficient FL as an example of how a survey can be useful despite imperfect execution. Send it to a serious referee with a clear instruction to verify the screening numbers and the taxonomy assignments. With those fixes, it becomes a legitimate survey; without them, it remains an overstated one.","headline":"A useful but rough entry-point survey of one-shot FL whose 'comprehensive reference' framing outruns its execution; worth a serious referee, not a desk reject.","tokens_in":794,"tokens_out":749,"would_cite":false,"duration_ms":24434,"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":"One-shot federated learning can train a global model in a single round, and this survey maps the field's methods, gaps, and next steps.","keywords":["Federated Learning","One-shot Federated Learning","Communication efficiency","Knowledge distillation","non-IID data","Model aggregation","Privacy preservation","Deep Learning"],"falsifier":"Re-running the paper's stated search across the three databases and checking every included paper against the four taxonomy categories would settle the completeness claim; a result in which many core one-shot FL methods fall outside all four categories would falsify the taxonomy. Similarly, locating the announced Section 4 on open-source resources and finding it absent from the text would falsify the claim to be a comprehensive reference.","tokens_in":22448,"feed_emoji":"📡","tokens_out":7302,"duration_ms":82785,"temperature":0.7,"pith_summary":"This survey argues that one-shot federated learning—training a global model in a single communication round instead of many—is a realistic answer to the communication bottleneck that limits standard federated learning. It organizes the field into four method families: theory-based optimization, knowledge-distillation ensembles, data-heterogeneity handling, and adversarial robustness. It also collects the datasets and open-source implementations used across the literature and outlines open challenges in non-IID data, trustworthiness, scalability, and the privacy-accuracy trade-off. A sympathetic reader should come away with a usable map of which methods exist, when each applies, and where the field most needs new work.","feed_headline":"One-shot federated learning cuts training to a single round","feed_subtitle":"A survey organizes methods, datasets, and open problems into a practical roadmap for the field.","key_machinery":"The load-bearing mechanism is the two-stage server-side pipeline: ensemble aggregation followed by synthetic-data distillation. A generator produces synthetic examples whose labels are chosen uniformly, and the ensemble's soft predictions supervise the final server model through KL divergence, so the quality of the synthetic set and the ensemble determine final accuracy. All four taxonomy categories are defined relative to this pipeline: theory-based methods refine the optimization objective, distillation methods improve the ensemble or the synthetic data, heterogeneity methods correct for non-IID gaps, and adversarial-robustness methods protect the pipeline.","core_discovery":"The paper's central claim is that one-shot FL is a coherent subfield with a shared working mechanism: clients train locally once, the server builds an ensemble of their logits, and a final global model is produced by distilling that ensemble on server-generated synthetic data. The survey formalizes this as minimizing cross-entropy on the combined private data, computing the ensemble as a weighted sum of client logits, and minimizing KL divergence between the ensemble and the server model. It then claims that the field's progress can be captured by four categories and that the principal unresolved barriers are data heterogeneity, privacy and security, scalability, and privacy-accuracy trade-offs, with near-term applications in IoT, satellite constellations, healthcare, and scientific machine learning.","pith_inferences":["A testable extension is that one-shot FL accuracy under extreme non-IID data should track the fidelity of server-generated synthetic data more strongly than ensemble size, since distillation transfers only what the synthetic set represents.","The privacy-accuracy costs reported in the survey suggest that one-shot FL may need privacy-aware synthetic-data generation, such as differentially private generators, rather than applying noise directly to the single model exchange.","The taxonomy's four categories are likely not disjoint; dataset-distillation methods straddle the distillation and heterogeneity categories, so a future map may need a cross-cutting axis for data-generation strategy."],"forward_implications":["If the survey's picture is right, one-shot FL cuts communication from $O(mn)$ exchanges to $O(m)$, which can make collaborative training feasible on bandwidth-limited and energy-constrained clients.","The four-category taxonomy gives practitioners a direct selection criterion: theory-based methods when guarantees matter, distillation when no public data is allowed, heterogeneity-aware methods for non-IID silos, and robustness methods under attack.","The paper's open-challenge analysis implies that near-term accuracy gains will come from better server-side synthetic-data generation and privacy-aware aggregation, not from adding communication rounds.","Adopting the proposed benchmark metrics—accuracy, communication cost, privacy loss, and scalability—would make one-shot FL results comparable across papers for the first time."],"supporting_citations":[{"why":"Introduces the one-shot FL problem and motivates single-round communication as the field's defining constraint.","marker":"Guha et al. (2019)"},{"why":"Supplies the theoretical limits for one-shot FL that the survey's theory-based category builds on.","marker":"Salehkaleybar et al. (2021)"},{"why":"FedKT is the canonical knowledge-transfer method the survey uses as a distillation and differential-privacy baseline.","marker":"Li et al. (2020a)"},{"why":"DENSE provides the data-free generator-plus-distillation template central to the synthetic-data pipeline.","marker":"Zhang et al. (2022)"},{"why":"FEDCVAE addresses very high statistical heterogeneity, a load-bearing example for the data-heterogeneity category.","marker":"Heinbaugh et al. (2023)"},{"why":"FedOV targets comprehensive label skew and anchors the distillation-based approach to non-IID data.","marker":"Diao et al. (2023)"},{"why":"FedFisher uses Fisher information to give theoretical guarantees, grounding the theory-based optimization family.","marker":"Jhunjhunwala et al. (2024)"},{"why":"XorMixFL's XOR mixup is an early privacy-preserving augmentation method for one-shot non-IID settings.","marker":"Shin et al. (2020)"},{"why":"FEN's two-stage aggregator bridges one-shot and standard FL and supports model heterogeneity, informing the scalability discussion.","marker":"Allouah et al. (2024)"}],"fun_headline_variants":["One-shot FL: one round, zero iterative communication","One-round federated learning: a survey of advances and challenges","Train once, aggregate once: the one-shot FL survey","One-shot federated learning: cutting communication to a single round","Single-round federated collaboration: one-shot FL survey"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey is useful only if its literature search was complete and its category assignments accurate; if it missed a substantial share of relevant work, or if the promised open-source section is missing, its taxonomy and recommendations will mislead readers.","fun_headline_variants_meta":{"raw":{"variants":["One-shot FL: one round, zero iterative communication","One-round federated learning: a survey of advances and challenges","Train once, aggregate once: the one-shot FL survey","One-shot federated learning: cutting communication to a single round","Single-round federated collaboration: one-shot FL survey"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000369,"raw_usage":{"total_tokens":1922,"prompt_tokens":831,"completion_tokens":1091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":1010}},"tokens_in":447,"tokens_out":1091,"duration_ms":12753,"temperature":1.0,"reasoning_tokens":1010,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:51:11.786508+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the paper's stated search across the three databases and checking every included paper against the four taxonomy categories would settle the completeness claim; a result in which many core one-shot FL methods fall outside all four categories would falsify the taxonomy. Similarly, locating the announced Section 4 on open-source resources and finding it absent from the text would falsify the claim to be a comprehensive reference.","supporting_citations":[{"cited_title":", author Sharifnassab, A","cited_arxiv_id":null,"evidence_quote":"Supplies the theoretical limits for one-shot FL that the survey's theory-based category builds on."},{"cited_title":", author Chen, C","cited_arxiv_id":null,"evidence_quote":"DENSE provides the data-free generator-plus-distillation template central to the synthetic-data pipeline."},{"cited_title":", author Wang, S","cited_arxiv_id":null,"evidence_quote":"FedFisher uses Fisher information to give theoretical guarantees, grounding the theory-based optimization family."}],"review_version":1}