{"id":"a043ee22-10e6-4041-9dd1-8af790a5008d","arxiv_id":"2412.19979","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"The paper combines semantic communication, federated learning with partial model freezing guided by Fisher information, and a leaky-ReLU activation-mapping heatmap to train and explain fire-detection models on industrial edge devices.","lead":"The paper proposes a federated learning system for fire surveillance in which industrial cameras send only compact semantic features instead of full images, plus a heatmap tool to explain the fire alerts. A reader might care because the framework targets three practical problems at once: cutting wireless bandwidth, keeping sensor data private, and making automated fire warnings interpretable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The delay-dominance claim rests on the assumption that freezing 1-ζ_n of a model's weights cuts local training time proportionally (Sec. IV), but forward/backward FLOPs do not scale linearly with trainable parameters, so XSFL's shortest-delay and accuracy-delay dominance are not established.","rationale":"The reader identified the Fisher-information freezing premise as the weakest assumption, focusing on the accuracy side of the claim. I identify a distinct but closely related unsupported premise on the delay side: Section IV's complexity analysis treats training cost as proportional to the number of trainable parameters, but Eq. (11) defines local delay in terms of the full local model, and forward computation through frozen layers is not eliminated. Since XSFL does not reduce communication size, its delay advantage over compression-based baselines depends entirely on this questionable computation-saving assumption. This is load-bearing because the strongest claim is simultaneous dominance in both accuracy and delay. I do not allege the method cannot work; the issue is that the empirical delay claim has no supporting timing evidence. The concrete test would settle it by measuring actual scaling. If measured delay reduction is near-proportional to the frozen fraction, the concern is resolved; if not, the Section V-C conclusion should be weakened. This keeps the reader's CONDITIONAL verdict unchanged: accept only after the timing experiment and, ideally, a reproducible artifact.","tokens_in":12222,"tokens_out":8079,"duration_ms":91283,"concrete_test":"Implement the XSFL ACT strategy on the same SC architecture and CDD split used in Fig. 3, and measure per-round wall-clock training time per device for frozen fractions 0, 0.25, 0.5, and 0.75, using both random masks and the Eq. (18) Fisher-selected masks, on the Section V-A CPU/channel settings, over at least 5 random seeds. Compare measured delay scaling to the linear-in-trainable-parameters assumption in Section IV and to FedAvg, STC, FTTQ, and FedPAQ under identical hyperparameters. If the measured reduction is substantially less than the frozen fraction, or if XSFL is not the fastest, the accuracy-delay dominance claim in Sections V-B/V-C must be revised or removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central delay-dominance claim rests on a computation-cost assumption that Section IV does not justify. Section IV states that XSFL's computation time complexity is O(T G Davg |\\hat w^{n,tr}_{g,t}|), i.e., linear in the number of trainable parameters, because ACT freezes 1−ζ_n of each device's weights. But Eq. (11) defines the local training delay as D_n Φ_n(w_n,t) G / f_n,t, where Φ_n is the cost of a forward-backward pass of the local model w_n,t. In a DNN, freezing a weight does not remove its forward multiplication: a frozen layer still contributes to forward activations and, if earlier layers remain trainable, backward gradient propagation still passes through it. Consequently, actual training-time savings from partial freezing are typically far smaller than the frozen fraction. Since XSFL does not reduce the transmitted model size (Section IV itself notes |\\hat w^n_{g,t}| = |w_{n,t}|), the only delay advantage over FedAvg, STC, FTTQ, and FedPAQ is this uncertain computation saving. Without wall-clock timing or FLOP counts, the Section V-C claim that XSFL gives the shortest average training delay, and hence the accuracy-delay dominance in the strongest claim, is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Industrial Edge Semantic Network (IESN) for fire surveillance, in which IIoT devices transmit only semantic features to an edge server. To train the semantic communication (SC) model while preserving privacy, the authors introduce XSFL (eXplainable Semantic Federated Learning), which combines an Adaptive Client Training (ACT) strategy and an Explainable SC (ESC) mechanism. ACT clusters devices by local data volume and freezes a device-specific fraction of global model weights selected via a Fisher-information-based importance measure, with the aim of reducing local training delay and improving accuracy. ESC uses a leakyReLU-based activation mapping to produce heatmaps that visualize which input regions drive the extracted semantics. The paper reports experiments on the Comprehensive Disaster Dataset (CDD) with ten simulated IIoT devices, comparing XSFL against FedAvg, STC, FTTQ, and FedPAQ, and claims that XSFL achieves the best accuracy/precision/specificity/F1/recall and the shortest average training delay, hence the best accuracy-delay trade-off.","tokens_in":12558,"tokens_out":3321,"duration_ms":36751,"significance":"If the central claims were substantiated, XSFL would be a useful contribution: it addresses a real problem (privacy-preserving, resource-adaptive, explainable semantic communication for industrial edge fire surveillance) and combines ideas from federated learning, Fisher-information-based weight freezing, and saliency-style explanation in a novel way. The paper also gives a clear system architecture and a reasonably organized set of algorithms. However, the evidence provided is currently too weak to support the claimed dominance over the baselines. The performance claim rests on a single dataset, ten simulated devices, no error bars or significance tests, no code, and no ablation of the ACT freezing criterion. The delay advantage is derived from an assumption that freezing parameters reduces per-sample computation cost linearly, which is not justified for deep neural networks. The explainability evaluation is qualitative only. These gaps are load-bearing because the paper's main selling points are exactly the accuracy-delay trade-off and the explainability mechanism.","major_comments":[{"comment":"The claim that XSFL has the shortest training delay is not established. Section IV states that the computation time complexity of XSFL is O(T G Davg |\\hat w^{n,tr}_{g,t}|), i.e., linear in the number of trainable parameters, but Eq. (11) defines the local training delay as D_n Φ_n(w_n,t) G / f_n,t, where Φ_n(w_n,t) is the cost of a forward-backward pass of the full local model. In a deep neural network, freezing a weight does not remove its forward multiplication or necessarily remove it from backward gradient propagation, so the actual computation savings are typically much smaller than the frozen fraction. Since Section IV itself notes that |\\hat w^n_{g,t}| = |w_{n,t}|, the communication complexity is unchanged, and the only claimed advantage over the baselines is this uncertain computation saving. The paper needs wall-clock training-time measurements or FLOP counts, not just a complexity expression, to support the Section V-C claim of shortest average delay and the optimization-objective comparison in Fig. 5.","section":"Section IV and Eq. (11)"},{"comment":"The accuracy-dominance claim is supported only by one dataset (CDD), ten simulated IIoT devices, and radar plots without error bars, confidence intervals, or significance tests. The simulation settings omit essential details: the SC model architecture, the number of communication rounds T, the local epochs G, the learning rate and optimizer, the number of clusters C in ACT, the actual ζ_n values per device, the LeakyReLU slope α in Eq. (25), and the way the Fisher information matrix in Eq. (18) is computed. The paper's own explanation for the accuracy gain is speculative (\"We speculate the excellent performance of XSFL is attributed to the ACT strategy\"). Without ablations that vary the freezing criterion, the freezing fraction, and the cluster count, the claim that XSFL achieves the best performance in all indicators is not demonstrated.","section":"Section V-A and V-B"},{"comment":"The ACT strategy's core premise is not validated. The Fisher information matrix I_n(w_n,t) is introduced without a definition, and Eq. (21) claims that I_n^g,t approximates (1/2)(w_g,t - w_n,t)^T H_n,t (w_g,t - w_n,t) at w_n,t = w_g,t, but the relationship between the quadratic form in Eq. (18) and the Hessian of the local loss is not derived. More importantly, the paper provides no evidence that freezing the weights with the smallest values of I_n^g,t improves local adaptation or global generalization; an ablation comparing Fisher-based freezing with random freezing or magnitude-based freezing is needed. As written, the only delay benefit of ACT is an artifact of the computation model in Eq. (11), so the accuracy benefit is an empirical claim that currently lacks supporting experiments.","section":"Section III-B, Eqs. (18)-(21)"},{"comment":"The statement that ESC \"can be used ... without any pre-training, hence there is no extra computation cost\" is internally inconsistent with Eqs. (22)-(24), which require computing first, second, and third derivatives of the semantic features with respect to feature-map activations. These gradient computations have real computational cost and may be expensive on resource-constrained IIoT devices. The paper should either remove the no-extra-cost claim or provide a complexity analysis and empirical timing for the ESC heatmap generation.","section":"Section III-C, Algorithm 3"}],"minor_comments":[{"comment":"The notation Mean(C_i^n) in Eq. (17) is ambiguous: C_i^n is defined as a set, but it is not clear whether it contains data volumes D_n or device indices; the text should specify that the mean is taken over the data volumes of devices in the cluster.","section":"Section III-B, Eq. (17)"},{"comment":"Radar plots without numerical values overlaid are hard to read; please provide a table with exact values and standard deviations across multiple random seeds.","section":"Section V-B, Fig. 3"},{"comment":"The figures lack clear axis labels and units; specify what \"training delay\" measures (wall-clock time or simulated CPU cycles) and what the optimization-objective weights are in Fig. 5 beyond the stated coefficient 1e-3.","section":"Section V-C, Figs. 4 and 5"},{"comment":"All devices are set to the same CPU frequency f_n,t = 2 GHz in the experiments, so the heterogeneity in computational resources that motivates ACT is not actually simulated; please vary f_n,t and D_n to demonstrate the claimed adaptivity.","section":"Section II-D, Eq. (11)"},{"comment":"The summation index in Eq. (25) uses k both as the kernel index and as the upper limit; please use a different variable, e.g., sum over e=1 to K, and make the definition of K explicit.","section":"Section III-C, Eq. (25)"},{"comment":"The explainability evaluation is purely qualitative; consider adding a quantitative metric such as insertion/deletion scores or pointing-game accuracy to substantiate the claim that ESC \"accurately marks\" fire and smoke regions.","section":"Section V-D"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant topic and has a reasonable structure, but the central claims are currently undersupported. The delay-dominance argument is tied to a computation-cost assumption that is not valid for modern DNN training, and the accuracy-dominance claim lacks statistical rigor and ablations. These issues are fixable within the scope of a major revision, provided the authors add wall-clock or FLOP-based timing, multi-seed experiments with error bars, ablation studies for ACT, and complete hyperparameter disclosure. I do not see grounds for rejection, but the current version should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: the XSFL paper is a coherent system paper that combines semantic communication, federated learning, and explainability for fire surveillance, but its central empirical claims are not yet supported. The delay advantage in particular is probably overstated.\n\nThe genuinely new piece is the ACT strategy: using a Fisher-weighted global-local distance to decide which weights to freeze per client, with the frozen share set by local data volume. That is a sensible idea and the writing around it is clear. The ESC heatmap is a small leaky-ReLU variant of Grad-CAM, but it is properly described and easy to implement. The authors also honestly note that communication complexity is unchanged. The system-level motivation is real: bandwidth, privacy, heterogeneity, and explainability all matter for this application.\n\nThe soft spots are substantial. Section IV claims computation time scales with the number of trainable parameters, so freezing a fraction of weights cuts training delay proportionally. That is not how DNN training works: forward propagation still passes through frozen layers, and backward passes still propagate through them if earlier layers are trainable. The real speedup from partial freezing is typically much smaller than the frozen fraction. Without wall-clock measurements or FLOP counts, the Section V-C result that XSFL has the shortest training delay is not established. The accuracy claim is also thin: one dataset, ten simulated devices, no error bars, no significance tests, no code. The key hyperparameters (local epochs, rounds, learning rate, cluster count, Fisher computation) are missing, and the paper itself says 'we speculate' that ACT causes the accuracy gain. An ablation that changes the freezing criterion would be the minimum needed.\n\nThat said, the paper is not confused or dishonest. It just tries to do too much with too little experimental support. The right outcome is major revision: measure actual compute, run ablations on the freezing rule, and release code and data.\n\nThis is a paper for people working on SC/FL applications who want a high-level integration template. It does not break new methodological ground, but it deserves a serious referee rather than a desk reject. I would engage with it on the condition that the experiments are rebuilt around real timing and ablations.\n\nRecommendation: send to review, but expect heavy revision.","headline":"A coherent SC+FL+explainability system paper whose delay advantage rests on a shaky compute model and whose accuracy claims lack experimental support.","tokens_in":13067,"tokens_out":3338,"would_cite":false,"duration_ms":30286,"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":"This paper aims to establish that a semantic federated learning framework, XSFL, can make industrial fire surveillance spectrally efficient, privacy-preserving, and explainable by freezing per-device weights chosen by Fisher information…","keywords":["semantic communication","federated learning","fire surveillance","explainable AI","model freezing","Fisher information","industrial Internet of Things","edge computing"],"falsifier":"A decisive experiment is to run XSFL on the CDD dataset with the same frozen-weight fractions, but choose the frozen weights by a random index set instead of by the Fisher-information score. If random freezing reproduces XSFL's reported accuracy, precision, F1, and delay, then the importance-based selection rule is not what carries the result; if random freezing clearly underperforms, the Fisher criterion is doing the claimed work.","tokens_in":12018,"feed_emoji":"🔥","tokens_out":7755,"duration_ms":75122,"temperature":0.7,"pith_summary":"This paper aims to establish that fire surveillance over an industrial edge network can be made spectrally efficient, privacy-preserving, and explainable in one framework. The framework, eXplainable Semantic Federated Learning (XSFL), lets IIoT cameras transmit only compact semantic symbols instead of raw frames, and trains the semantic encoder-decoder cooperatively at the edge so raw data never leaves the devices. To handle heterogeneous devices, the Adaptive Client Training (ACT) strategy freezes a per-device fraction of model weights chosen by a Fisher-information importance score, which is reported to cut training delay while preserving or improving accuracy. An explainable semantic communication (ESC) mechanism then produces heatmaps that align extracted semantics with fire and smoke regions in the input image. On the CDD fire dataset, the paper reports that XSFL beats FedAvg, STC, FTTQ, and FedPAQ on accuracy, precision, specificity, F1-score, and recall, and that it gives the shortest average training delay per device.","feed_headline":"Federated learning with frozen weights speeds fire alerts","feed_subtitle":"Reported on fire-image benchmarks, it beats four baselines on accuracy while cutting per-device training delay.","key_machinery":"The load-bearing mechanism is the Adaptive Client Training (ACT) strategy. For device $n$, ACT clusters devices by local data volume, sets a trainable proportion $\\zeta_n = \\mathrm{Mean}(C_n^i)/D_{\\max}$, and freezes the $1-\\zeta_n$ fraction of global weights with the smallest values of the Fisher-weighted quadratic $I_{n}^{g,t} = (w_{g,t} - w_{n,t})^\\top \\mathrm{diag}(I_n(w_{n,t}))(w_{g,t} - w_{n,t})$, using the Fisher information matrix as a Hessian approximation. Freezing low-importance weights is what reduces local computation per device while allowing the remaining weights to adapt to local data. The second mechanism is the Explainable SC (ESC) mechanism, which computes gradients of semantic features with respect to the last convolutional feature map, weights them with second-order coefficients, and applies leaky ReLU to form a localization heatmap; this is what connects a prediction back to image regions.","core_discovery":"The central discovery is that federated training of a semantic communication model can be adapted to heterogeneous IIoT devices by freezing different weight subsets per device, and that this simultaneously improves the accuracy-delay trade-off and preserves explainability. Concretely, the paper claims XSFL achieves the best values on all five classification indicators on the CDD fire dataset, and that its average training delay per device is the shortest among the compared schemes. The reported reason is that ACT's weight freezing, guided by the Fisher information of each local model, lets limited-resource devices update fewer parameters while well-resourced devices train more, improving local adaptation and global generalization. The ESC heatmaps, the paper argues, show that the extracted semantics correspond to visible fire and smoke regions, making the predictions interpretable.","pith_inferences":["A test not run in the paper: compare ACT's Fisher-based freezing against random freezing of the same fraction on the same dataset; if accuracy ties, the Fisher criterion itself is not the driver of the reported gain.","The explainability evaluation in the paper is qualitative; a quantitative faithfulness score for the heatmaps, such as deleting highlighted regions and measuring the accuracy drop, would let future work compare ESC against other explanation methods.","If the reported five-metric dominance holds under varying channel noise, the same framework could extend to other disaster-surveillance tasks such as smoke, flood, or industrial anomaly alerts, but that transfer is not established here."],"forward_implications":["Fire-surveillance devices could send semantic symbols rather than raw frames, reducing uplink bandwidth while retaining classification accuracy.","The maximum per-round delay in synchronous federated learning shrinks because slower devices train fewer weights, so more devices can stay within the delay budget.","The ESC heatmaps give rescue crews a human-checkable link between a fire alert and the image evidence, which is the kind of transparency needed for life-critical decisions.","The reported optimization objective, combining global loss and weighted delay, gives a single number for comparing federated schemes on the accuracy-latency trade-off."],"supporting_citations":[{"why":"Serves as the FedAvg baseline that XSFL must beat in accuracy and delay comparisons.","marker":"[11]"},{"why":"Provides the k-medoids clustering method used to group IIoT devices by local data volume, which sets each device's freeze proportion.","marker":"[19]"},{"why":"Supplies the Fisher-information-to-Hessian approximation that justifies freezing weights with low importance scores.","marker":"[20]"},{"why":"Gives the gradient-weighted activation mapping approach that the ESC mechanism adapts for semantic heatmaps.","marker":"[24]"},{"why":"Provides the CDD fire dataset used for all training and evaluation in the paper.","marker":"[27]"},{"why":"Supplies the STC compression-based federated learning baseline compared in the simulations.","marker":"[28]"},{"why":"Supplies the FTTQ ternary-compression baseline compared in the simulations.","marker":"[29]"},{"why":"Supplies the FedPAQ periodic-averaging and quantization baseline compared in the simulations.","marker":"[30]"}],"fun_headline_variants":["Adaptive weight freezing boosts federated fire alerts","Explainable federated learning for fast fire detection","Federated learning adapts to devices for fire warnings","Fisher-based freezing speeds federated fire monitoring","Semantic federated learning yields explainable fire heatmaps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's accuracy advantage rests on the premise that freezing the $1-\\zeta_n$ fraction of weights with the smallest Fisher-information-based $I_{n}^{g,t}$ values improves local adaptation and global generalization without degrading accuracy; the premise is asserted through the Hessian approximation, and no ablation varies the freezing criterion.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive weight freezing boosts federated fire alerts","Explainable federated learning for fast fire detection","Federated learning adapts to devices for fire warnings","Fisher-based freezing speeds federated fire monitoring","Semantic federated learning yields explainable fire heatmaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1203,"prompt_tokens":866,"completion_tokens":337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":482,"tokens_out":337,"duration_ms":3952,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:42:29.372727+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive experiment is to run XSFL on the CDD dataset with the same frozen-weight fractions, but choose the frozen weights by a random index set instead of by the Fisher-information score. If random freezing reproduces XSFL's reported accuracy, precision, F1, and delay, then the importance-based selection rule is not what carries the result; if random freezing clearly underperforms, the Fisher criterion is doing the claimed work.","supporting_citations":[{"cited_title":"Analysis of k-means and k-medoids algorithm for big data,","cited_arxiv_id":null,"evidence_quote":"Provides the k-medoids clustering method used to group IIoT devices by local data volume, which sets each device's freeze proportion."},{"cited_title":"A comparative study on disaster detection from social media images using deep learning,","cited_arxiv_id":null,"evidence_quote":"Provides the CDD fire dataset used for all training and evaluation in the paper."},{"cited_title":"Ternary compression for communication-efficient federated learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the FTTQ ternary-compression baseline compared in the simulations."}],"review_version":1}