{"id":"f68134f2-c62f-49a6-8a26-dfa3266345fd","arxiv_id":"2608.11532","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A hierarchical federated transfer learning scheme for vehicular digital twins reports higher accuracy and efficiency than federated baselines, but its evaluation lacks a fair pre-training control and error bars.","lead":"The paper proposes a hierarchical federated transfer learning framework for digital twin-based vehicular networks, grouping vehicles by type, fine-tuning pre-trained models per cluster, and weighting updates with blockchain-based trust scores. The experiments claim improvements over centralized, federated, and clustered federated baselines, but the comparison may be unfair because only the proposed method is given a pre-trained model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evaluation confounds transfer learning with the proposed HFTL algorithm: only HFTL starts from a pre-trained model, so the reported gains do not isolate the hierarchical clustering, digital twin, or reputation mechanisms.","rationale":"The reader's weakest assumption precisely identifies the load-bearing flaw: HFTL is the only method that starts from a pre-trained model. The manuscript itself confirms this in Section 3.1 and in the discussion of the results, which attributes faster training, lower resource use, and higher accuracy to fine-tuning. Because no ablation separates the pre-trained initialization from the proposed hierarchical clustering, digital twin layer, or reputation-based weighting, the central empirical claim is unsupported. The absence of error bars, the lack of released code, the untested trust mechanism, and the unfinished template placeholders are additional concerns, but they are secondary to the confounded baseline. The appropriate verdict is unchanged: reject as submitted, with the path to acceptance being controlled experiments that isolate the algorithm's contribution.","tokens_in":14537,"tokens_out":2320,"duration_ms":26683,"concrete_test":"Add two control conditions to the Section 4 experiments: (a) 'HFTL-scratch', which runs the same hierarchical clustering, intra-cluster relaying, and trust-weighting procedure but initializes the model randomly instead of using a pre-trained checkpoint; (b) pre-trained baselines, in which FL, CFL, and CL are initialized from the same pre-trained model used by HFTL. Keep the dataset split, number of epochs (200), learning rate (0.1), batch size (128), and 100 simulation runs identical, and report mean and standard deviation for every metric in Table 3. If HFTL still outperforms the pre-trained baselines, the hierarchical mechanism adds value; if HFTL-scratch matches or outperforms pre-trained HFTL, or if pre-trained baselines close the gap, the reported advantage is due to transfer learning rather than the proposed architecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that HFTL delivers higher accuracy, lower training time, lower resource consumption, lower latency, and faster convergence than CL, FL, and CFL in DT-VANETs. The evidence for this claim is Table 3 and Figure 4, backed by the scalability runs in Table 4 and Figure 6. However, the baseline comparison does not isolate the proposed mechanism. Section 3.1 states that HFTL fine-tunes a pre-trained model shared by vehicle type, while Section 4.1.4 describes training the baseline algorithms with the same hyperparameters but gives no indication that CL, FL, or CFL receive any pre-trained initialization. The experimental write-up repeatedly attributes the advantage to this asymmetry: fine-tuning requires less time than training from scratch, pre-trained models require less time for predictions, and customized pre-trained models improve accuracy. In other words, the measured superiority of HFTL is exactly what transfer learning alone would predict. The additional HFTL-specific components—clustering by vehicle type, head/route/edge relaying, digital twin synchronization, and reputation-weighted aggregation—are never ablated. The trustworthiness mechanism is described but not evaluated in any experiment, so the only quantitative evidence for the central claim is the confounded comparison. As submitted, the results do not establish that the proposed hierarchical federated transfer learning architecture contributes anything beyond a warm-started model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hierarchical Federated Transfer Learning (HFTL) for Digital Twin-based Vehicular Ad Hoc Networks. Vehicles are clustered by type; a pre-trained model is fine-tuned inside each cluster by head, route, and edge vehicles; and updates are weighted by a data-quality reputation score recorded on a blockchain. The authors present two algorithms (Weighted Cloud Server Cycling Model Update and Inner Cluster Federated Transfer Learning) and evaluate them on a real vehicle mobility trace against Centralized Learning, Federated Learning, and Clustered Federated Learning. Tables 3 and 4 and Figures 4-7 report that HFTL achieves higher accuracy, lower training time, lower resource consumption, lower latency, faster convergence, and higher throughput, at the cost of slightly higher communication overhead.","tokens_in":14794,"tokens_out":5714,"duration_ms":55332,"significance":"If the claimed results were properly supported, HFTL would be a meaningful contribution to personalized, privacy-preserving model training in vehicular networks: per-vehicle-type fine-tuning combined with reputation-based aggregation is a sensible design, and the scalability tests address a practically important regime. The paper also considers multiple performance metrics and uses a real-world mobility trace. However, the current evaluation does not separate the effect of transfer learning from the proposed clustering, digital twin, and trust mechanisms; the reputation system is never experimentally tested; and no uncertainty quantification is provided. The core idea is promising, but the evidence presented is not sufficient to support the central claim.","major_comments":[{"comment":"The central empirical claim rests on a confounded comparison. Section 3.1 states that HFTL fine-tunes a pre-trained model shared according to vehicle type, while Section 4.1.4 describes the baseline algorithms (CL, FL, CFL) as trained with the same hyperparameters but gives no indication that any baseline receives a pre-trained initialization. The discussion in Section 4.1.6 explicitly attributes the advantages to this asymmetry (e.g., 'fine-tuning a model requires less time compared to training a model from scratch' and 'pre-trained models that require less time to make predictions'). Consequently, Tables 3 and 4 and Figures 4 and 6 do not isolate the contributions of clustering, digital twin synchronization, or reputation weighting; the improved accuracy and convergence could be produced by transfer learning alone. Please add a control in which FL and CFL are initialized from the same pre-trained model, and ablate the HFTL-specific components (clustering by type, hierarchical relay, reputation weighting) one at a time.","section":"Section 4.1.4 vs. Section 3.1"},{"comment":"The trustworthiness and reputation mechanism is described but never evaluated. Equations (2)-(7) define the data quality score, reputation update, model-update weight, and exclusion threshold, and Section 3.4 claims that this mechanism keeps malicious vehicles from interfering; however, no experiment in Section 4 varies the presence of malicious vehicles, compares HFTL with and without reputation weighting, or tests the exclusion threshold of Equation (7). The abstract and contribution list therefore claim a security property that the experimental section does not measure.","section":"Section 3.4 and Section 4"},{"comment":"The performance comparison lacks uncertainty quantification. The text states that 100 simulations were run to eliminate bias, yet Table 3 and Figure 4 report only average values with no standard deviation, confidence interval, or per-simulation spread. Without such information, the reader cannot assess whether the differences between HFTL (model accuracy 0.822) and CFL (0.752) are stable across the random topologies of Figure 3. Please report the distribution of outcomes.","section":"Table 3 and Figure 4"}],"minor_comments":[{"comment":"The notation is inconsistent: Equation (2) defines DTS_z, but Equation (3) uses DQS_z; Equation (3) defines RPS_z(t_m), while Equation (5) writes RS_z(t_m). Please unify the symbols.","section":"Equations (2)-(5)"},{"comment":"Equation (1) is not a well-formed formal definition; the expression 'DT = ΣΣ Tvj∈Ci' is ambiguous. Please define DT as a set or collection of clusters and state the indexing explicitly.","section":"Equation (1)"},{"comment":"Equation (8) introduces the notation PSr and |v| without formal definitions in the surrounding text; also, the function Fweight(...) used in Algorithm 1 is not defined beyond this equation.","section":"Equation (8)"},{"comment":"Section 4.1.2 says there are twelve random clusters each with more than ten vehicles, while Section 4.1.3 says vehicles are 'randomly clustering vehicles according to their similar category'; please clarify whether clusters are homogeneous by type or whether the type assignment is post-hoc.","section":"Section 4.1.2"},{"comment":"The manuscript contains unresolved placeholders: 'Research highlight 1' and 'Research highlight 2' after the highlights header, and the label 'Graphical Abstract' before the title. These should be completed or removed.","section":"Highlights and Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The main technical issue is the missing pre-training control, and the fix is a set of additional experiments; if the authors cannot supply them in revision, the paper should not be accepted. The manuscript also seems under-polished (placeholders, notation inconsistencies, unrelated citations), so the editor may want to apply stricter presentation standards."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nYou should know two things about this one. The architecture is coherent and the application is sensible; the experiments, however, do not test the architecture. Only the proposed method gets a pre-trained model, so the reported gains in accuracy, training time, and convergence are what transfer learning alone would predict. The hierarchical clustering, digital twin synchronization, and reputation weighting are never isolated or ablated.\n\nWhat the paper does well: the idea of clustering vehicles by type and fine-tuning a shared pre-trained model per cluster is a reasonable combination of existing techniques. The authors describe two algorithms (cloud server cycling update and inner-cluster transfer) with enough detail to reproduce the mechanics. They test on a real mobility trace, report scalability results, and are honest about limitations (data heterogeneity, network instability, model bias). That is more than many submissions do.\n\nThe soft spots are hard to miss. Section 3.1 says HFTL fine-tunes a pre-trained model; Section 4.1.4 gives the same hyperparameters to baselines but never says they receive any pre-trained initialization. The text then repeatedly attributes HFTL's advantages to fine-tuning versus training from scratch—which is exactly the confound. There is no error bar or variance measure despite 100 simulation runs, no ablation of the trust or clustering components, and the blockchain-based reputation mechanism is described but never evaluated. Notation is inconsistent (DTS vs DQS, RS vs RPS), and the manuscript still contains template placeholders and off-topic references (e.g., antibody-related Ahmed et al. citations).\n\nThe central claim—that HFTL outperforms CL, FL, and CFL—is not established by the evidence as presented. That said, the problem is a real one and the fix is clear: repeat the comparison with all methods pre-trained (or none), add ablations, report error bars, and actually implement the trust layer. With those changes the paper could become a modest contribution to vehicular federated learning.\n\nWho is this for? Researchers working on hierarchical federated transfer learning in ITS may find the architecture a useful starting point, but not the empirical claims. I'd bring it to a reading group as a case study in confounded experimental design.\n\nMy recommendation: don't accept as is. If a resubmission corrects the baselines and adds the missing controls, it deserves peer review. For this round, I'd lean toward major revision rather than outright rejection, but the placeholders make it hard to argue it's fully ready for referees.\n\nBest,\n[You]","headline":"Sensible architecture, but the evaluation gives only HFTL a pre-trained model, so the reported gains are transfer learning by construction; the distinct contributions are unvalidated.","tokens_in":15360,"tokens_out":2944,"would_cite":false,"duration_ms":29971,"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":"Grouping vehicles by type and fine-tuning pre-trained models yields faster, more accurate federated learning in digital-twin vehicular networks.","keywords":["hierarchical federated transfer learning","digital twin","vehicular ad hoc network","federated learning","vehicle clustering","trustworthiness metrics","blockchain","model personalization"],"falsifier":"Retrain the three baselines under exactly the same settings as HFTL but give them the same pre-trained initialization, then compare accuracy and convergence time. If federated learning or clustered federated learning with pre-training matches HFTL's 0.822 accuracy and 0.04s convergence, the proposed hierarchy is not the source of the improvement and the central claim is not supported.","tokens_in":14273,"feed_emoji":"🚗","tokens_out":7627,"duration_ms":70391,"temperature":0.7,"pith_summary":"This paper proposes a training scheme called Hierarchical Federated Transfer Learning (HFTL) for digital-twin vehicular networks. Vehicles are grouped by type—regular cars, emergency vehicles, buses, delivery vehicles, specialized vehicles, motorcycles—and each group fine-tunes a shared pre-trained model on its own private local data. Updates travel hierarchically from edge vehicles through route vehicles to a head vehicle and then to the cloud, where they are weighted by blockchain-recorded data-quality and reputation scores and averaged into the global model. The authors claim this produces higher model accuracy, lower training time, lower resource consumption, lower latency, and faster convergence than centralized learning, federated learning, and clustered federated learning on a real vehicle mobility trace, with only a small increase in communication overhead. If correct, the scheme would give each vehicle type a dedicated, accurate predictive model without sharing raw sensor data.","feed_headline":"Vehicle-type model fine-tuning beats federated learning","feed_subtitle":"HFTL clusters vehicles by type, fine-tunes shared models, and weights updates by trust scores, lifting accuracy and cutting training time.","key_machinery":"The load-bearing mechanism is a two-level aggregation protocol wrapped around pre-trained-model fine-tuning. Algorithm 1 (Weighted Cloud Server Cycling Model Update) has the cloud cycle through the collection of vehicle-type clusters, send the current model to each cluster in turn, and update the global model only when the cluster's average reputation score clears a threshold. Algorithm 2 (Inner Cluster Federated Transfer Learning) organizes each cluster as a chain: a head vehicle receives the pre-trained parameters from the base station, forwards them to route vehicles, which pass them to edge vehicles; each vehicle fine-tunes with its private data and returns the update along with its data quantity and quality score, and the head combines updates weighted by data quantity and score before sending the result back up. The reputation update rule and the threshold check are what keep malicious or low-quality participants out of the global update.","core_discovery":"The claim at the center of the paper is that the combination of vehicle-type clustering, pre-trained-model fine-tuning, and trust-weighted hierarchical aggregation is what makes federated learning work in heterogeneous vehicular settings. Because each cluster starts from a model already trained on a related task and refines it on data from vehicles of one type, the resulting model matches that type's driving patterns and data distribution instead of being diluted by averaging across very different vehicles. The trustworthiness metric—built from data completeness, sensor collaboration, event reporting, format consistency, vehicle health, and safe driving—decides how much weight a vehicle's update gets and can exclude a vehicle entirely once its reputation score falls below a threshold. In the reported experiments, HFTL reaches 0.822 average accuracy versus 0.752 for clustered federated learning, 0.721 for federated learning, and 0.708 for centralized learning, while reducing training time to 1.02s and convergence time to 0.04s; communication overhead is slightly higher than the baselines. The paper interprets these results as evidence that HFTL is well suited to real-time ITS applications like traffic-flow prediction, route optimization, predictive maintenance, and collision alerts.","pith_inferences":["An implication the paper leaves implicit: the same cluster-and-fine-tune structure should transfer to other digital-twin settings with behaviorally distinct device classes—industrial fleets, drones, or robots—provided a related pre-trained source model exists.","A natural ablation would isolate the contribution of each design choice: compare HFTL against a version with random (non-type) clustering, a version without reputation weighting, and a version trained from scratch, to see which component actually drives the accuracy and convergence gains.","Because the data-quality score is computed by the vehicle itself before being stored on the blockchain, a malicious vehicle could in principle inflate its own score; in practice the reputation mechanism would need independent verification of the score components to close that hole.","The measured tradeoff suggests a testable boundary: with model compression or asynchronous updates, the communication overhead could drop below standard federated learning while the accuracy and convergence advantages remain, which would make the case for deployment stronger."],"forward_implications":["A deployed HFTL system would maintain a distinct predictive model for each vehicle category rather than one global model, so emergency, delivery, and public-transport vehicles each get predictions tuned to their own routes and data.","Because each cluster fine-tunes a pre-trained model, training time and computational resource use drop relative to training from scratch, easing the load on vehicles with limited onboard compute.","The reputation threshold gives the network a concrete defense against malicious participants: a vehicle whose score falls below the threshold cannot contribute to the global model, and the blockchain prevents post-hoc tampering with those scores.","The scalability experiments indicate that as the network grows to hundreds of vehicles, HFTL keeps accuracy near 0.82 while the baseline methods fall to roughly 0.68, so the relative benefit widens at larger scale.","The slight communication overhead increase is the price paid for transmitting pre-trained models and personalized updates, and the paper identifies communication compression and asynchronous updates as the levers for reducing it in very large deployments."],"supporting_citations":[{"why":"Introduces federated learning, the base training paradigm that HFTL modifies and compares against.","marker":"Konečný et al. (2016)"},{"why":"Introduces federated transfer learning, the pre-trained-model knowledge-transfer mechanism HFTL builds upon.","marker":"Liu et al. (2020)"},{"why":"Prior hierarchical federated learning in multi-hop cluster-based VANETs; supplies the head/route/edge structure and the clustered federated learning comparison.","marker":"HaghighiFard and Coleri (2024)"},{"why":"Positions federated learning within digital twin-based vehicular networks and motivates the architecture's application context.","marker":"Khan et al. (2023)"},{"why":"Documents security and privacy challenges in vehicular digital twin networks, motivating the blockchain-backed trust-score design.","marker":"He et al. (2022)"},{"why":"Compares transfer learning, split learning, and federated learning for ITS, providing the comparative framing for federated transfer learning in vehicular settings.","marker":"Otoum et al. (2022)"}],"fun_headline_variants":["Vehicle-type clustering lifts federated learning accuracy","Trust-weighted federated transfer learning for vehicular nets","Hierarchical transfer learning beats vanilla FL in DT-VANET","Cluster by vehicle type, fine-tune, and score trust for FL","HFTL: vehicle-type clustering improves federated learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes the reported gains come from the hierarchical clustering and trust weighting, but only HFTL starts from a pre-trained model while the baselines train from scratch, so the advantage could be mostly the known benefit of fine-tuning rather than the proposed architecture.","fun_headline_variants_meta":{"raw":{"variants":["Vehicle-type clustering lifts federated learning accuracy","Trust-weighted federated transfer learning for vehicular nets","Hierarchical transfer learning beats vanilla FL in DT-VANET","Cluster by vehicle type, fine-tune, and score trust for FL","HFTL: vehicle-type clustering improves federated learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1533,"prompt_tokens":962,"completion_tokens":571,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":490}},"tokens_in":578,"tokens_out":571,"duration_ms":5853,"temperature":1.0,"reasoning_tokens":490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:36:22.744870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the three baselines under exactly the same settings as HFTL but give them the same pre-trained initialization, then compare accuracy and convergence time. If federated learning or clustered federated learning with pre-training matches HFTL's 0.822 accuracy and 0.04s convergence, the proposed hierarchy is not the source of the improvement and the central claim is not supported.","supporting_citations":[],"review_version":1}