{"id":"ced6c37a-d713-4d45-86a0-9fd95745ccdf","arxiv_id":"2505.04873","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A survey of federated learning for cyber physical systems, covering architectures, applications, challenges, and future directions, with a proposed integration framework.","lead":"This paper reviews how federated learning is applied to cyber physical systems, including smart healthcare, smart cities, vehicles, and security. It organizes the field into taxonomies, outlines a general integration framework, and lists future research directions.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No search protocol or inclusion criteria is presented, so the 'first comprehensive and systematic analysis' claim is unsupported; several quantitative entries in the taxonomy tables are also not traceable to the cited papers.","rationale":"The reader's weakest assumption correctly identifies the absence of a systematic selection method as the central problem with the survey's headline claim. I agree that this is load-bearing: 'comprehensive' and 'systematic' are explicit in Section I.B, and without a reproducible protocol the paper cannot support that claim. I additionally flag a closely related weakness that the reader mentioned in the rationale but did not make the headline weakest assumption: the taxonomy tables contain specific quantitative results and limitation attributions that are not verifiable from the cited sources. Even a perfectly exhaustive reference list would not fix inaccurate or untraceable table entries, so the survey's value as a reference depends on both conditions. The paper does have genuine strengths: it covers many application domains, provides extensive taxonomy tables, and discusses challenges and future directions, so a complete rejection would be too harsh. However, the central claim should be presented conditionally, with the methodology made explicit or the claim softened. The reader's CONDITIONAL verdict is appropriate, so no verdict change is needed.","tokens_in":44613,"tokens_out":5204,"duration_ms":56223,"concrete_test":"Perform a systematic literature audit: run a PRISMA-style search across Scopus, Web of Science, and IEEE Xplore for (\"federated learning\" AND (\"cyber-physical system\" OR \"CPS\")) over 2019-2025, deduplicate, screen with explicit inclusion criteria matching the paper's stated scope, and compare the final list with the references in Tables VI-VIII and X-XI. If any substantial cluster of relevant FL-CPS papers is missing from any application domain, the 'comprehensive' claim fails. In the same audit, randomly sample 10 quantitative cells from these tables and verify each against the full text of the cited paper; an unlocatable number or limitation statement would require a table correction and would further undermine the survey's reference value.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section I.B, bullets 1 and 3) is that this paper provides the 'first comprehensive and systematic analysis' of FL-CPS and taxonomy tables that classify the 'technical components, contributions, and limitations' of prior work. For that claim to hold, the set of included papers must be selected in a reproducible way, and the synthesized statements about each paper must be accurate. Neither condition is established. Section II describes FL and CPS foundations but gives no search protocol, no inclusion/exclusion criteria, no database list, and no quality assessment; the survey therefore cannot distinguish 'comprehensive' from 'representative but arbitrary.' The problem is not merely procedural: Section V.C explicitly concedes that there is 'no standard approach for comparing' FL-CPS solutions and that existing works are tested on 'distinct network setups and data, making direct comparison challenging,' yet the tables (VI, VII, VIII, X, XI) present cross-paper numbers as comparable (e.g., '95.45% accuracy' for [127], 'MAE: 12.8 kW' for [166], 'F1=0.93' for [210]) without a common benchmark. Several limitation cells, such as 'Requires TEEs' in Table VI for reference [134], are not traceable to statements in the cited articles. If these entries cannot be confirmed, the survey's own stated purpose as a reference for researchers and practitioners is materially weakened.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of federated learning (FL) for cyber-physical systems (CPS). It begins with background on FL and CPS, then proposes an FL-CPS taxonomy and integration framework, reviews applications in healthcare, smart cities, vehicular systems, and cybersecurity, and concludes with lessons learned, open challenges, and future directions. The paper's central claim (Section I.B) is that it provides the 'first comprehensive and systematic analysis' of FL-CPS integration, along with taxonomy tables that classify technical components, contributions, and limitations of prior work.","tokens_in":44922,"tokens_out":3421,"duration_ms":31481,"significance":"If the central claim were substantiated, this survey would be a valuable reference for the growing FL-CPS community. The paper covers a broad and relevant set of application domains, and its organization into taxonomy tables and a proposed integration framework is a reasonable structure for a survey. The authors are to be credited for assembling a large body of work and for identifying several important research gaps. However, the significance is currently limited by the lack of a transparent literature selection methodology and by the poor traceability of many quantitative claims in the taxonomy tables. Because the paper positions itself as a definitive reference, these issues affect the core value of the contribution and must be addressed before the survey can serve its intended purpose.","major_comments":[{"comment":"The paper claims to provide the 'first comprehensive and systematic analysis' of FL integration within CPS (Section I.B, bullets 1 and 3), but no literature search protocol, inclusion/exclusion criteria, database list, or quality assessment is described in Section II or elsewhere. Without such methodology, the survey cannot distinguish a comprehensive and reproducible selection of papers from a representative but arbitrary one. This directly undermines the paper's central contribution. The authors should either add a detailed methodology subsection describing how papers were identified, screened, and quality-assessed, or temper the claim of 'comprehensive and systematic' to 'broad and representative.'","section":"Section I.B and Section II"},{"comment":"The taxonomy tables contain numerous quantitative entries (e.g., '95.45% accuracy' for [127], 'MAE: 12.8 kW' for [166], 'F1=0.93' for [210], 'Requires TEEs' in Table VI for [134]) that are not traceable to the cited papers. The 'Limitations' column in particular includes assertions such as 'no formal convergence guarantees for non-IID data' and 'Requires TEEs for secure aggregation' that are either absent from or not clearly supported by the cited sources. Since the paper's stated purpose is to serve as a reference for researchers and practitioners (Section I.B), the accuracy and provenance of every taxonomy entry is load-bearing. The authors should either provide a direct citation or derivation for each quantitative claim and limitation, or remove entries that cannot be verified.","section":"Tables VI, VII, VIII, X, XI"},{"comment":"Internal cross-referencing is broken in several places: Section IV.C.2.b contains the fragment 'predicting traffic patterns IV-C2a,' which appears to be an incomplete reference; Section IV.A.2.b refers to 'Section IV .B's focus on real-time health monitoring' when it means the remote health monitoring subsection of the same section; and Table IV's 'Taxonomy Reference' column points to 'Section III.B.3,' 'Section III.B.2,' and 'Section III.B.1,' but Section III.B in the manuscript has no such numbered subsections. Furthermore, the survey structure shown in Fig. 1 does not match the actual section numbering in the text (e.g., Fig. 1 shows Section III as 'FL-CPS Taxonomy Framework' and Section IV as 'FL-CPS Applications,' which does align, but internal references such as 'Section III.B.3' do not correspond to any visible structure). These errors make the survey difficult to use and should be systematically corrected.","section":"Section IV.C.2.b and Fig. 1"},{"comment":"Section V.C explicitly states that there is 'no standard approach for comparing' FL-CPS solutions and that existing works are tested on 'distinct network setups and data, making direct comparison challenging.' Yet the taxonomy tables present cross-paper numbers as if they were directly comparable, without any qualifier about differing datasets, hardware, or evaluation protocols. This internal tension undermines the reliability of the tables as a reference tool. The authors should either add a prominent caveat that all numbers are as reported by the original papers under heterogeneous conditions, or restructure the tables so that the metrics are presented as unverified literature values rather than as a meaningful comparative benchmark.","section":"Section V.C and quantitative comparisons"}],"minor_comments":[{"comment":"There are numerous typos and inconsistent spellings, including 'deice' for 'device' (Section I.A), 'UA Vs' and 'UAV' inconsistencies, 'addtion' for 'addition' (Section IV.B.2.d), 'beyong-5G' for 'beyond-5G' (Section IV.D.1), and 'RestNet50' for 'ResNet50' (Section IV.B.2.a). A careful proofreading pass is needed.","section":"Throughout"},{"comment":"Reference [157] is cited for 'fuzzified one-way hashes' but does not appear in the visible reference list, and the citation numbering appears to be out of order in several places (e.g., [157] before [158] in the text). The reference list should be checked for completeness and correct ordering.","section":"Section IV.A.2.d"},{"comment":"The proposed hierarchical aggregation formula in Eq. (3) introduces a free parameter α (set to 0.7 in the text) without any justification or sensitivity analysis. Since this is presented as part of a general integration framework, the choice of α should be explained or explicitly marked as an example rather than a default.","section":"Section II.C, Eq. (3)"},{"comment":"Table IV's 'Taxonomy Reference' column uses section numbers that do not exist in the manuscript (e.g., 'Section III.B.3'). This should be corrected to point to the actual subsections of the taxonomy framework or removed if the framework is self-contained.","section":"Section II.C, Table IV"}],"recommendation":"major_revision","confidential_remarks":"The large number of unverifiable quantitative entries in the taxonomy tables is a serious concern for a survey that claims to be a definitive reference. During the revision, the authors should be asked to provide a source for every numeric claim and every limitation statement, or to remove those entries. If this cannot be done, the paper's credibility as a survey will be materially weakened. The lack of a literature selection methodology also needs to be addressed, either through added transparency or through a more modest positioning of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuinely useful mapping of the FL-CPS literature, but it does not deliver the \"first comprehensive and systematic analysis\" it promises. The scope is ambitious—healthcare, smart cities, vehicular systems, and cybersecurity—and the taxonomy tables give newcomers a structured entry point. The IoT-versus-CPS comparison is a helpful framing that most prior surveys lack. I also think the proposed five-layer integration framework (data, model, aggregation, control, verification) is a reasonable organizing device, though it is descriptive, not validated.\n\nWhat the paper does well is synthesize a large body of work into categories that make sense: architecture types (centralized, hierarchical, decentralized, hybrid), data characteristics, learning paradigms, and privacy tiers. That structure will help a graduate student find relevant papers quickly. The authors have read a lot and the citation list is wide.\n\nThe soft spots are real. The stress-test note lands: there is no search protocol, no inclusion criteria, no quality assessment, so \"systematic\" is a stretch. Worse, the quantitative entries in the taxonomy tables appear without any common benchmark or, in some cases, without support in the cited text. For example, Table VI attributes \"Requires trusted execution environments (TEEs)\" to [134], but the section discussing [134] says nothing about TEEs. Several accuracy numbers (95.45%, F1=0.93, MAE 12.8 kW) are given as if they were directly comparable, even though Section V.C explicitly concedes that existing works use \"distinct network setups and data.\" That is a traceability problem, not a matter of taste. The framework section also has some un-sourced numbers (e.g., 10-20% communication reduction, alpha=0.7 in Eq. 3) that need citations or removal.\n\nThere are also internal slips: typos like \"deice\" for device, and some cross-references that point to the wrong subsection. These are minor, but they add to the impression of a manuscript that was not carefully proofread.\n\nWho gets value from this? Someone new to FL-CPS who wants a bird's-eye view of application areas and a starting taxonomy. I would not use it as a reliable reference for specific numbers without checking the primary sources. The survey is a useful secondary source, not a definitive one.\n\nI would send it to peer review, because the scope and framing are valuable and the authors are credible. But I would make the review contingent on a transparent methodology section, removal or verification of unsupported quantitative table entries, and a clear disclaimer that this is a structured survey rather than a systematic review.","headline":"Useful mapping of FL-CPS with a helpful taxonomy, but the 'systematic' claim is not supported and several table numbers don't check out; deserves peer review only with major revisions.","tokens_in":45405,"tokens_out":3487,"would_cite":false,"duration_ms":34564,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A survey organizes federated learning for cyber-physical systems into a taxonomy and a five-layer integration framework.","keywords":["federated learning","cyber-physical systems","machine learning","privacy protection","taxonomy","smart healthcare","intelligent transportation","smart cities"],"falsifier":"Run a reproducible literature search for federated learning in cyber-physical systems over the same period and check whether a substantial share of the retrieved papers can be placed in the taxonomy's four architecture categories and five framework layers; if many cannot, the claim of comprehensive coverage fails. A single search for a prior peer-reviewed survey with the same FL-CPS scope would also falsify the 'first comprehensive analysis' claim.","tokens_in":44466,"feed_emoji":"🤖","tokens_out":7252,"duration_ms":71115,"temperature":0.7,"pith_summary":"This paper assembles and organizes the scattered research on federated learning in cyber-physical systems. It argues that FL's privacy-preserving, decentralized training protocol fits the constraints of CPS—distributed sensors, real-time decision making, safety, heterogeneous devices, and data sovereignty—and that the field has matured enough to be mapped. The paper builds a taxonomy of FL-CPS designs and a five-layer integration framework, then applies them to healthcare, smart cities, vehicular systems, and cybersecurity. If the map is right, researchers and practitioners gain a shared vocabulary, a way to compare systems, and a checklist of the open problems that still block practical deployment.","feed_headline":"A survey maps federated learning for cyber-physical systems","feed_subtitle":"The taxonomy and five-layer framework give designers a shared structure and a list of open problems.","key_machinery":"The load-bearing object is the FL-CPS taxonomy and integration framework. The taxonomy sorts systems along four dimensions: architecture types (centralized, hierarchical, decentralized, and hybrid), data characteristics (non-IID distribution, volume, variety, and velocity), learning paradigms (model-centric, data-centric, and hybrid), and privacy requirements (basic encryption, differential privacy with homomorphic encryption, and secure multi-party computation with zero-knowledge proofs). The integration framework organizes the lifecycle into a data layer, a model layer, an aggregation layer, a control layer, and a verification layer, linked by a five-phase workflow spanning distributed data acquisition, model initialization, collaborative training, secure aggregation, and deployment with feedback. These structures do the argument's work: they turn a scattered literature into a coordinate system for positioning individual systems and for naming what is still missing.","core_discovery":"The paper's central claim is that federated learning and cyber-physical systems have converged into a recognizable field that can be systematically described for the first time. It asserts that FL's distributed, privacy-preserving training model is a natural fit for CPS, where data live on heterogeneous sensors and actuators, decisions must respect real-time safety constraints, and raw data cannot be centralized. To make this case, the paper constructs a taxonomy with four classification dimensions (architecture type, data characteristics, learning paradigm, and privacy requirement) and a five-layer integration framework (data, model, aggregation, control, and verification), then applies both to four application clusters: healthcare, smart cities, vehicular systems, and core cybersecurity. The claim also includes a comparative analysis that separates CPS from IoT and a roadmap of open problems—security, resource management, standardization, heterogeneity, and communication costs—that the paper says must be solved before FL-CPS can be broadly deployed.","pith_inferences":["Beyond the paper: the taxonomy could be used as a benchmark grid, and one testable prediction is that hybrid architectures will dominate deployments where both real-time edge autonomy and global retraining matter.","Beyond the paper: the five-layer framework suggests a maturity model—a new FL-CPS system could be scored layer by layer, turning a descriptive map into an engineering checklist.","Beyond the paper: the three privacy tiers imply a quantitative cost curve; an experiment running the same FL-CPS task under all three tiers would reveal the accuracy and latency price of stronger privacy guarantees.","Beyond the paper: because the survey finds no standard specification for federated CPS, a natural test is whether two independently built FL-CPS testbeds can federate with each other today; if they cannot, standardization is the binding constraint."],"forward_implications":["FL-CPS systems can now be described and compared along common axes—architecture type, data characteristics, learning paradigm, and privacy tier—so a new deployment can be positioned against existing work.","The five-layer integration framework gives system builders a reusable structure for moving from sensor-level data collection to verified, safety-checked model deployment.","The survey's gap analysis points to concrete priorities: robust aggregation against poisoning and backdoor attacks, energy-aware client selection, formal verification of safety constraints, and standards for comparing federated CPS deployments.","The IoT versus CPS comparison implies that FL solutions designed for the IoT cannot be assumed to work in CPS without accounting for real-time control loops and safety certification.","The application surveys in healthcare, smart cities, vehicular networks, and cybersecurity show that the same FL mechanisms—FedAvg, FedProx, split learning, secure aggregation—transfer across domains."],"supporting_citations":[{"why":"Supplies the prior comprehensive survey of FL for IoT that this paper positions itself against and extends to CPS.","marker":"[8]"},{"why":"Provides the FedAvg algorithm and the client–server workflow that the paper treats as the canonical FL mechanism.","marker":"[9]"},{"why":"Serves as a comparison baseline showing that earlier FL surveys do not cover CPS-specific challenges.","marker":"[17]"},{"why":"Another comparison baseline used to establish the gap this survey fills.","marker":"[18]"},{"why":"Documents FL for resource-constrained IoT devices, a related line the survey distinguishes from CPS demands.","marker":"[20]"},{"why":"Covers FL in mobile edge networks, a neighboring application area used to position the CPS scope.","marker":"[24]"},{"why":"Provides the smart-healthcare FL survey that anchors the healthcare CPS section.","marker":"[25]"},{"why":"Reviews CPS for manufacturing and supplies the CPS-specific framing and open challenges.","marker":"[31]"}],"fun_headline_variants":["Mapped: federated learning for cyber-physical systems","Federated learning meets CPS: a comprehensive map","Taxonomy and framework for federated learning in CPS","How federated learning is reshaping cyber-physical systems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey assumes that the papers it selected, without a documented search protocol or inclusion criteria, are representative enough to make its taxonomy and gap analysis complete.","fun_headline_variants_meta":{"raw":{"variants":["Mapped: federated learning for cyber-physical systems","Federated learning meets CPS: a comprehensive map","Taxonomy and framework for federated learning in CPS","How federated learning is reshaping cyber-physical systems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000702,"raw_usage":{"total_tokens":3183,"prompt_tokens":978,"completion_tokens":2205,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":2141}},"tokens_in":594,"tokens_out":2205,"duration_ms":14952,"temperature":1.0,"reasoning_tokens":2141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:18:15.371053+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a reproducible literature search for federated learning in cyber-physical systems over the same period and check whether a substantial share of the retrieved papers can be placed in the taxonomy's four architecture categories and five framework layers; if many cannot, the claim of comprehensive coverage fails. A single search for a prior peer-reviewed survey with the same FL-CPS scope would also falsify the 'first comprehensive analysis' claim.","supporting_citations":[],"review_version":1}