{"id":"7909b3a9-fb7b-4212-b02b-196a0a9d2bb2","arxiv_id":"2501.01063","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper sketches an untested architecture integrating adaptive privacy, dynamic masking, blockchain provenance, secure aggregation, and XAI for IoV federated learning.","lead":"This paper proposes an architecture that combines adaptive privacy, dynamic masking, blockchain, secure aggregation, and explainable AI for federated learning in the Internet of Vehicles. It does not implement, measure, or formally verify any of its claimed security and scalability properties.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The framework's security and scalability claim hinges on the unvalidated assumption (Section IV-B) that blockchain can perform real-time, high-throughput logging and consensus at IoV scale; without a protocol or benchmark, the claimed guarantees are unsupported.","rationale":"I concur with the reader's REJECT verdict. The paper is an architecture proposal with no implementation, simulation, or formal correctness proof. The most load-bearing weakness is the Section IV-B blockchain scalability assumption: all security claims filter through the blockchain logging and validation stage, but the paper gives no protocol parameters or performance evidence to make this stage plausible at IoV scale. The BAN logic section is not a proof—it is a list of assumptions and desired beliefs with no derivation. Table I is explicitly subjective. That said, the paper deserves credit for explicitly stating assumptions and for referencing relevant prior work (e.g., SHERPA, IFed, FL-DECO-BC); this makes it a plausible starting point for a systems paper. However, an architecture whose core guarantees depend on an untested scalability premise cannot support the abstract's assertion that it 'assures' security, scalability, and interpretability. The concrete simulation test described above would settle whether the premise can be met.","tokens_in":10926,"tokens_out":3555,"duration_ms":34346,"concrete_test":"Build a minimal simulation of Steps 2 and 3: N vehicle nodes (e.g., 100, 1k, 10k) each submit one masked gradient per FL round (round duration D seconds) to a permissioned Proof-of-Stake blockchain; measure end-to-end validation latency and accepted-updates-per-second. Compare achieved throughput against the required rate N/D and against published throughput of baseline systems (e.g., PoS Ethereum, Hyperledger Fabric). If the accepted-update rate is below N/D or latency exceeds the real-time bound, the scalability assumption fails and the architecture's security guarantees do not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FAPL-DM-BC is secure, scalable, and interpretable. Section IV-B explicitly assumes the blockchain \"is scalable enough to perform high-frequency, high-throughput operations including logging, validation, and consensus in real-time.\" This is load-bearing because every subsequent security property—tamper-proof logging (Step 2), provenance, replay and man-in-the-middle resistance—depends on each model update being validated and recorded before aggregation (Step 3). The paper provides no consensus design, throughput or latency analysis, nor any benchmark; it even concedes in Section IV-B that this is an assumption. If a fleet's update rate exceeds blockchain capacity, updates are delayed or dropped, and the system must either discard them or fall back to centralized aggregation, both of which breach the claimed guarantees. The BAN logic section does not rescue this: it restates beliefs and goals but derives no theorem, and the comparison table (Table I) carries an explicit disclaimer that the analysis was \"interpreted subjectively.\" Thus the architecture's core promises rest on an untested premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FAPL-DM-BC, a federated learning architecture for the Internet of Vehicles that combines adaptive privacy (FAPL), dynamic masking, blockchain-based secure logging and provenance, SMPC-based secure aggregation with FedAvg, and a dual-model XAI feedback loop. The authors describe a six-step workflow, list assumptions, provide a BAN logic 'analysis', and compare the framework with prior work in a table. The central claims are that the framework is secure (anonymity, non-traceability, resistance to replay, man-in-the-middle, message modification, and other attacks), scalable, and interpretable. No experiments, simulations, formal proofs, or system measurements are reported.","tokens_in":11075,"tokens_out":5359,"duration_ms":46688,"significance":"If the claims were validated, the framework could be a useful integration of known techniques for privacy-preserving and auditable federated learning in vehicular networks. The paper is clearly written in its component descriptions and is honest about its main assumption concerning blockchain scalability. However, as it stands the significance is limited: the security and scalability properties are asserted rather than demonstrated, the only formal-looking component (BAN logic) does not contain an actual derivation, and the comparison table is explicitly subjective. There are no falsifiable experimental outcomes, no reproducible code, and no quantitative analysis, so the paper does not meet the evidentiary standard for a security or systems contribution.","major_comments":[{"comment":"The framework assumes that the blockchain infrastructure is scalable enough to perform high-frequency, high-throughput operations including logging, validation, and consensus in real-time. This assumption is load-bearing: Steps 2-4 of the workflow in Section IV-A require every model update to be logged, validated, and committed on the blockchain before aggregation, and the security properties in Section V (non-traceability, replay resistance, tamper-proof provenance) all depend on that pipeline. The paper provides no consensus protocol design, no throughput or latency analysis, and no benchmark to substantiate this assumption. The stated assumption therefore does the work of the missing evaluation.","section":"Section IV-B, first paragraph"},{"comment":"The security claims are presented as assertions without a threat model or verification. For example, Section V-A states that anonymity is achieved through differential privacy noise and gradient masking, but the paper never defines the privacy mechanism's parameters (e.g., an epsilon-delta bound), nor does it analyze re-identification or membership-inference risk. Similarly, Section V-D asserts replay and man-in-the-middle resistance based on timestamps and encryption, but no protocol specification or adversarial analysis is provided.","section":"Section V, including V-A to V-D"},{"comment":"The BAN logic analysis does not establish the stated goals. It lists assumptions and 'logical steps' as prose, but it never applies BAN inference rules (e.g., jurisdiction, nonce-verification, or message-meaning rules) to derive the goals from the assumptions. For instance, the goal P |≡ C |∼ MP is not derived; Step 1 only asserts C |≡ P |∼ MP. Consequently, the passage does not amount to a formal verification.","section":"Section IV-C"},{"comment":"The comparative evaluation in Table I is not a valid basis for claiming superiority. The note states that the entries were 'interpreted subjectively' where sources did not explicitly provide details, and the table gives no methodology, metrics, or per-cell citations. The claim that FAPL-DM-BC outperforms previous frameworks on 18 attack categories is therefore unsupported.","section":"Table I"},{"comment":"The 'weighted mean computation' Final Updates = wL · x + wG · y is not justified as a mechanism. The paper does not specify how wL and wG are computed, does not impose a normalization constraint (so the expression is not necessarily a weighted mean), and provides no convergence or accuracy analysis for the resulting update rule. Because the abstract credits this equation with assuring secure, scalable, and interpretable FL, it is load-bearing and needs substantive support.","section":"Section IV-A6, Eq. (1)"}],"minor_comments":[{"comment":"The abstract contains sentence fragments, e.g., 'Secure Logging and Verification, Blockchain-based provenance and decentralized validation, and Cloud Microservices Secure Aggregation using FedAvg ... and Secure Multi-Party Computation (SMPC).' This should be rewritten as complete sentences.","section":"Abstract"},{"comment":"The notation is inconsistent: the local feedback updates are called x in Eq. (1) but MP in the BAN logic section, and the global model updates are called y but MC. Please unify notation throughout the paper.","section":"Section IV-A6"},{"comment":"The phrase 'reducing the scope of false positives in the predictions' is undefined; no mechanism for false-positive reduction is described, so the claim should be clarified or removed.","section":"Section IV-A6"},{"comment":"Reference [6] describes spectral random masking for particle image velocimetry, not gradient masking in federated learning; this citation does not support the Dynamic Masking concept as used in the paper.","section":"References"},{"comment":"The manuscript contains numerous typographical and grammatical errors, such as 'cooccurrently', 'readability' for 'interpretability', 'impervious auditing', and 'globally knowable knowledge'; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper is an architecture proposal with no experimental validation; the central claims are unsupported. In my view it does not meet the bar for a security journal. Additionally, the proposed framework is closely related to the authors' prior FL-DECO-BC (ref [19]), and Section III and Table I position FAPL-DM-BC relative to that prior work; the novelty of FAPL-DM-BC beyond FL-DECO-BC should be scrutinized if the authors resubmit with evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clearly written architecture proposal, not a validated research result. The authors combine known FL privacy, masking, blockchain, SMPC, and XAI components into one IoV pipeline, describe a sensible workflow, and are upfront about their assumptions. That honesty is the paper's best feature. What it doesn't do is support the headline claims—secure, scalable, interpretable—with any implementation, simulation, benchmark, or formal proof.\n\nThe strongest section is the workflow description (IV-A): the sequence from local training with adaptive privacy, to blockchain logging, to SMPC aggregation, to XAI-based local feedback is coherent and readable. The BAN logic section (IV-C) is a reasonable attempt to frame trust goals, but it lists assumptions and steps rather than deriving conclusions, so it does not establish security. Table I is explicitly subjective; the disclaimer under it should be taken at face value.\n\nThe load-bearing problem, as flagged, is Section IV-B's assumption that the blockchain can handle high-frequency, high-throughput logging, validation, and consensus in real time. Every claimed guarantee—tamper-proof provenance, replay resistance, freshness—depends on that. No consensus protocol is specified beyond PoS, no throughput or latency numbers are given, and no benchmark is provided. If the fleet outruns the chain, updates get delayed or dropped, and the framework's guarantees break. The update rule 'Final Updates = wL·x + wG·y' is a definitional weighted mean; the weights are described qualitatively, not derived or fitted, so it contributes no quantitative content. The self-citation to FL-DECO-BC is natural given the comparison, but that table's subjectivity limits what it can support.\n\nWhat is genuinely new is narrow: a particular composition of existing components and a dual-model XAI feedback loop described at a high level. That may be useful as a starting design for a systems paper, but as submitted this is a position paper. It is not incoherent, and the authors are honest about assumptions—I would rather see this than a paper that hides them—but the central claims remain asserted.\n\nMy recommendation: do not send this to peer review as a research contribution. A serious referee would need an implementation or at least a realistic simulation measuring latency, throughput, and attack resistance. Ask the authors to come back with that. I would not cite it in my own work yet.","headline":"A well-organized architecture proposal with honest assumptions, but the security, scalability, and interpretability claims are asserted rather than demonstrated; no implementation, benchmark, or formal proof.","tokens_in":11688,"tokens_out":2554,"would_cite":false,"duration_ms":26317,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FAPL-DM-BC claims that federated learning for the Internet of Vehicles can be simultaneously privacy-adaptive, tamper-proof, scalable, and explainable through a weighted fusion of local XAI feedback and globally aggregated updates.","keywords":["Federated learning","Internet of Vehicles","Adaptive privacy","Dynamic masking","Blockchain","Explainable AI","Secure multiparty computation","IoV security"],"falsifier":"Run the framework on a simulated fleet of 1,000 vehicles and measure the latency of one federated round, isolating the blockchain logging and Proof-of-Stake validation step; if validation throughput cannot keep pace with the incoming update rate or end-to-end latency exceeds a real-time driving deadline, the central claim that the pipeline is scalable and real-time capable fails.","tokens_in":10682,"feed_emoji":"🚗","tokens_out":6746,"duration_ms":59679,"temperature":0.7,"pith_summary":"The paper proposes FAPL-DM-BC, a federated learning architecture for the Internet of Vehicles. Each vehicle trains locally, tunes privacy protections in real time (differential privacy or gradient masking) according to data sensitivity, and sends masked, encrypted updates through a blockchain-backed logging and validation layer before cloud microservices aggregate them with FedAvg and secure multi-party computation. A dual-model explainable-AI loop checks local predictions and explanations, and a weighted mean fuses that local feedback with the global model update. The intended result is a single framework that is simultaneously privacy-preserving, tamper-proof, scalable, and interpretable, which matters because autonomous driving and traffic management need all four at once.","feed_headline":"IoV federated learning gains adaptive privacy, blockchain, and XAI","feed_subtitle":"Vehicles train together without exposing raw data, with tamper-proof logging and explanations for every prediction.","key_machinery":"The central mechanism is the weighted-mean update rule $$\\text{Final Updates} = w_L \\cdot x + w_G \\cdot y,$$ where $x$ is the local feedback update produced by the dual-model XAI loop, $y$ is the global update from FedAvg, and $w_L$ and $w_G$ are weights reflecting local explanation quality and global dataset size and diversity. This rule is the point where the paper claims local interpretability and global knowledge reinforce each other rather than compete. Around this rule sit the supporting mechanisms: FAPL and Dynamic Masking adapt the privacy mechanism to data sensitivity, blockchain logs and validates each update through Proof of Stake, and SMPC keeps aggregation private.","core_discovery":"The paper's central claim is that the FAPL-DM-BC architecture combines four mechanisms, adaptive privacy tuning via Federated Adaptive Privacy-Aware Learning and Dynamic Masking, blockchain-based secure logging and decentralized validation, FedAvg plus Secure Multi-Party Computation for aggregation, and a dual-model XAI feedback loop, so that no single component has to be sacrificed to obtain the others. The framework asserts that every model update is hashed and verified before aggregation, that the aggregator never sees raw individual updates, and that local predictions are certified by an explainable second model before they influence the global model. If correct, the framework would cover every attack category in its comparison table while keeping the federated learning process transparent enough for safety-critical vehicular decisions.","pith_inferences":["If the framework works, the weighted-mean rule could be reused in other edge-federated settings, such as healthcare or smart-grid diagnostics, where local explanations should influence the global model.","The paper leaves the weights $w_L$ and $w_G$ as a design choice; a natural test is to measure accuracy, explanation fidelity, and poisoning resistance under different weighting schedules.","The dual-model validation loop could be extracted and deployed as a standalone local defense against poisoned updates before aggregation.","Because the security claims depend on blockchain scalability, a prototype on a real high-throughput blockchain would be the missing step that turns the architecture into an engineering result."],"forward_implications":["Vehicle edge nodes can switch privacy mechanisms on the fly as data sensitivity or threat level changes, instead of locking in a fixed privacy budget.","Every model update is hashed and recorded on the blockchain, so tampering after logging becomes detectable and provenance is auditable.","Global aggregation runs over masked and encrypted updates through SMPC, so the cloud aggregator does not see individual vehicle updates.","The dual-model XAI feedback loop lets each vehicle validate its own prediction and explanation before the update leaves the vehicle, which should reduce false positives.","If the comparison table is accurate, the framework claims coverage of every listed attack category, including front-running, traffic analysis, and location pinpointing, where some baseline frameworks are marked limited or absent."],"supporting_citations":[{"why":"Defines Federated Adaptive Privacy-Aware Learning, the adaptive privacy-tuning basis the framework builds on.","marker":"[5]"},{"why":"Supplies the Dynamic Masking technique used to obfuscate model updates in transit.","marker":"[6]"},{"why":"Provides the Secure Multi-Party Computation protocol used for privacy-preserving aggregation.","marker":"[7]"},{"why":"Establishes blockchain for IoV as the provenance and decentralized validation substrate.","marker":"[8]"},{"why":"Supplies the XAI methods (SHAP, LIME, Grad-CAM) used in the dual-model feedback loop.","marker":"[9]"},{"why":"Supplies the FedAvg aggregation rule used for global model updates.","marker":"[20]"},{"why":"Baseline framework combining XAI and federated learning against data poisoning; appears in the comparison table.","marker":"[15]"},{"why":"Baseline federated learning framework with local differential privacy; appears in the comparison table.","marker":"[16]"},{"why":"Prior blockchain-based federated learning framework for VANETs whose provenance approach this work extends and compares against.","marker":"[19]"}],"fun_headline_variants":["Adaptive privacy + blockchain + XAI: secure IoV federated learning","FAPL-DM-BC: adaptive privacy, blockchain, XAI for IoV federated learning","Secure IoV training: adaptive privacy, dynamic masking, blockchain, XAI","IoV federated learning gets transparent adaptive privacy and XAI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole secure aggregation and provenance pipeline depends on the blockchain handling high-frequency, high-throughput logging, validation, and consensus in real time; if a large fleet's update volume outruns the blockchain, the framework stalls.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive privacy + blockchain + XAI: secure IoV federated learning","FAPL-DM-BC: adaptive privacy, blockchain, XAI for IoV federated learning","Secure IoV training: adaptive privacy, dynamic masking, blockchain, XAI","IoV federated learning gets transparent adaptive privacy and XAI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000338,"raw_usage":{"total_tokens":1850,"prompt_tokens":909,"completion_tokens":941,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":855}},"tokens_in":525,"tokens_out":941,"duration_ms":8135,"temperature":1.0,"reasoning_tokens":855,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:35:11.575348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the framework on a simulated fleet of 1,000 vehicles and measure the latency of one federated round, isolating the blockchain logging and Proof-of-Stake validation step; if validation throughput cannot keep pace with the incoming update rate or end-to-end latency exceeds a real-time driving deadline, the central claim that the pipeline is scalable and real-time capable fails.","supporting_citations":[{"cited_title":"Privacy-aware federated learning for page recommendation,","cited_arxiv_id":null,"evidence_quote":"Defines Federated Adaptive Privacy-Aware Learning, the adaptive privacy-tuning basis the framework builds on."},{"cited_title":"Spectral random masking: A novel dynamic masking technique for PIV in multiphase flows,","cited_arxiv_id":null,"evidence_quote":"Supplies the Dynamic Masking technique used to obfuscate model updates in transit."},{"cited_title":"Secure multi-party computation,","cited_arxiv_id":null,"evidence_quote":"Provides the Secure Multi-Party Computation protocol used for privacy-preserving aggregation."},{"cited_title":"Blockchain for IoV in 6G environment: Review solutions and challenges,","cited_arxiv_id":null,"evidence_quote":"Establishes blockchain for IoV as the provenance and decentralized validation substrate."},{"cited_title":"Explainable AI (XAI): Core ideas, techniques, and solutions,","cited_arxiv_id":null,"evidence_quote":"Supplies the XAI methods (SHAP, LIME, Grad-CAM) used in the dual-model feedback loop."},{"cited_title":"Federated learning’s blessing: FedAvg has linear speedup,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg aggregation rule used for global model updates."},{"cited_title":"SHERPA: Ex- plainable Robust Algorithms for Privacy-Preserved Federated Learning in Future Networks to Defend Against Data Poisoning Attacks,","cited_arxiv_id":null,"evidence_quote":"Baseline framework combining XAI and federated learning against data poisoning; appears in the comparison table."},{"cited_title":"IFed: A novel federated learning framework for local differential privacy in Power Internet of Things,","cited_arxiv_id":null,"evidence_quote":"Baseline federated learning framework with local differential privacy; appears in the comparison table."},{"cited_title":"FL-DECO-BC: A Privacy-Preserving, Provably Secure, and Provenance-Preserving Federated Learning Framework with Decentralized Oracles on Blockchain for VANETs","cited_arxiv_id":"2407.21141","evidence_quote":"Prior blockchain-based federated learning framework for VANETs whose provenance approach this work extends and compares against."}],"review_version":1}