{"id":"fa2a205c-c480-434b-b315-bc4e6528478f","arxiv_id":"2611.03273","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"TIDE averages each sensor's encoder only among devices holding that sensor, and reports the best accuracy per communication budget on three activity recognition datasets.","lead":"TIDE is a federated learning method that averages each sensor's model only among the devices that carry that sensor, preventing devices without the sensor from erasing what the others learned. It reports the best accuracy per communication budget on three activity recognition datasets, including a 7.63 point gain over the next method on ActionSense.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The unhedged 'most accurate at B/2 on all three datasets' claim counts PFHAR reproduced from random initialization without its defining pretraining; the authors acknowledge the modification but never bound its effect, and PFHAR already posts 89.87 on UCI-HAR without it.","rationale":"The reader's weakest_assumption is that the comparison is fair, and I agree that this is where the argument is least secure; my read narrows it to a single decisive sub-case. What the central claim needs is that every baseline in Table I represents the method it is named after, at the communication budget of Eq. (3). That condition is least secure for PFHAR, because the paper's own setup sentence says it was run from random initialization without the large-scale pretraining that defines the method, and the direction of the induced error is known: removing pretraining can only lower early-budget accuracy, and the unhedged B/2 claims are small-margin claims (1.07 and 1.49 points, within reported seed variation on two of three datasets). PFHAR's no-pretraining Final on UCI-HAR (89.87) is already within 1.30 points of TIDE, so the upside is not hypothetical. I checked the other acknowledged modifications and found them less load-bearing: the approximate staged-communication conversion produces an anomalous FedHKD WEAR profile (3.46% at B vs 72.98% at termination), but the top rivals at B (RELIEF, FedMVD, MCARN) are non-staged, so that error does not affect the leading margins; the GraMFedDHAR noise-multiplier adjustment sits about 25 points below TIDE on its best dataset. The ActionSense headline, against a same-group non-staged baseline with full participation on both sides, is the component most likely to survive an independent check. One concrete run, PFHAR with its published pretrained initialization under the identical protocol, would settle whether the unhedged B/2 claim survives; until then CONDITIONAL remains the right verdict, so I do not change the reader's decision.","tokens_in":15711,"tokens_out":23941,"duration_ms":148672,"concrete_test":"Obtain PFHAR's published large-scale pretrained initialization and re-run the full Table I protocol (identical availability patterns, per-round communication metering, seeds 42/123/2024, early stopping) on UCI-HAR and WEAR, recomputing Final, @B, and @B/2. If PFHAR's @B/2 exceeds TIDE's 90.36 (UCI-HAR) or 80.07 (WEAR), the abstract's 'most accurate on all three datasets at half its own communication' fails against the unmodified baseline and must be re-hedged; otherwise the B/2 claim survives the strongest identified threat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is comparative, and its breadth rests on the unhedged sentence that TIDE is 'the most accurate on all three datasets at half its own communication' (abstract; conclusion repeats it). On UCI-HAR and WEAR the B/2 margins are 1.07 and 1.49 points over RELIEF and FedMVD, within the seed variation the paper itself reports, so the claim is settled only on ActionSense. The baseline field behind these small margins is produced under acknowledged modifications, the most consequential of which is PFHAR: Section IV-A states 'PFHAR is reproduced from random initialization without its large-scale pretraining, and hence its results reflect its fine-tuning alone.' The paper never bounds what that pretraining adds, and its removal can only lower PFHAR's early-budget accuracy, which is exactly what the B/2 comparison measures. Even without it, PFHAR reaches Final 89.87 on UCI-HAR, within 1.30 points of TIDE's 91.17, so a pretrained PFHAR could plausibly exceed TIDE's 90.36 and 80.07 at B/2 on UCI-HAR and WEAR, or even the ActionSense 78.06. The Table I footnote's approximate round-to-GiB conversion for staged Harmony and FedHKD is a genuine blemish (FedHKD's WEAR curve shows 3.46% at B = 41.0 GiB against 72.98% at termination), but it does not threaten the leading rivals at B, which are non-staged. The ActionSense margin over RELIEF (78.06 vs 70.43, non-staged, effectively full participation on both sides) is the most robust component; the load-bearing open question is the quantified impact of the pretrained PFHAR on the equal-budget claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper proposes TIDE, a federated-learning method for multimodal IoT sensing under two coupled forms of heterogeneity: devices carry different sensor subsets, and devices come online in periodic patterns. TIDE gives each sensor its own encoder and auxiliary classification head, aggregates those parameters only over the devices that hold the sensor ('intra-pod merging'), keeps the fusion layer and task head shared, augments local training with class prototypes, a cross-sensor alignment loss, and modality dropout, and selects clients by a weighted score of training speed, data size, and sensor-combination rarity. The evaluation is framed as accuracy within a communication budget B (Eq. (3)) and compares TIDE with eight baselines on UCI-HAR, ActionSense, and WEAR, reporting 78.06% on ActionSense (7.63 points above the next method), the highest mean accuracy at B/2 on all three datasets, and, at equal budget, a lead on ActionSense with the other two datasets within seed variation. An incremental ablation attributes most of the gain to the auxiliary heads and intra-pod merging, and the paper explicitly reports its limitations, including the failure of score-based selection at small K on WEAR.","tokens_in":16276,"tokens_out":23133,"duration_ms":135274,"significance":"The central mechanism is simple, well motivated, and falsifiable: only devices that hold a sensor should update its parameters, and the server knows exactly which devices those are. If the reported gains hold, intra-pod merging combined with per-sensor auxiliary heads is a useful addition to the multimodal-FL toolbox, and the paper improves on typical evaluation practice by comparing methods at equal communication (Eq. (3)) rather than at each method's own stopping point; this framing is what makes RELIEF's UCI-HAR lead interpretable (280.2 GiB of communication for a 1.42-point margin over TIDE's 75.2 GiB). The ablation reporting is honest: it shows that prototype alignment, the participation term, and cross-sensor alignment move accuracy within seed variation, that the auxiliary heads alone do not always help, and that score-based selection at small K trails random selection on WEAR, a limitation the authors state themselves. The ActionSense comparison (78.06% vs RELIEF's 70.43%, both at similar communication and effectively full participation) is the most robust component of the empirical case.","major_comments":[{"comment":"The unhedged claim that TIDE is 'the most accurate on all three datasets at half its own communication' is stated more strongly than the evidence supports on two of the three datasets. In Table I, the B/2 margins are 1.07 points on UCI-HAR (TIDE 90.36 vs RELIEF 89.29) and 1.49 points on WEAR (TIDE 80.07 vs FedMVD 78.58), while the paper's own three-seed standard deviations for Final are ±1.03 (TIDE, UCI-HAR) and ±0.77/±0.74 (TIDE/FedMVD, WEAR); on both datasets the margin is the same order of magnitude as the seed variation the abstract invokes in the equal-budget sentence ('ahead within seed variation on the other two'). Since the @B and @B/2 columns report no variance, the 'within seed variation' hedge cannot be checked for those columns. I recommend either hedging the B/2 claim in the same way as the equal-budget claim, or reporting per-seed values with a significance test; the ActionSense margin of 6.60 points over RELIEF at B/2 is robust and not in question.","section":"Abstract, Section IV-B, and Conclusion","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The B/2 comparison against PFHAR runs a modified method. Section IV-A states that 'PFHAR is reproduced from random initialization without its large-scale pretraining,' and the paper never bounds what that pretraining contributes; this is load-bearing for the headline B/2 claim. Pretraining can only be expected to raise a model's early-budget accuracy, which is exactly the quantity the B/2 comparison measures, and on UCI-HAR even the unpretrained PFHAR reaches a final 89.87%, within 1.30 points of TIDE's 91.17%. As reproduced, PFHAR is therefore not the method named by reference [32], and the claim that TIDE is the most accurate at B/2 on all three datasets is not conditioned on this modification. Please either quantify the effect (for example, report the published PFHAR results and the communication they require, or run a pretrained variant) or explicitly restrict the affected claims to the reproduced baseline field.","section":"Section IV-A and Table I","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The round-to-GiB conversion for the staged Harmony and FedHKD is acknowledged as approximate, but the conversion rule itself is not reported; on WEAR, FedHKD is listed at 3.46% for B = 41.0 GiB against a final 72.98% with Comm = 55.4 GiB, so the @B and @B/2 entries of the staged methods depend on an unreported quantity. Please state the per-round communication model or mark the affected entries as approximate in the table body.","section":"Table I footnote and Section IV-A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The selection weights reported in Section IV-A, α = 0.25, β = 0.35, and γ = 0.25, sum to 0.85 rather than 1; because the score in Eq. (5) is used only for ranking, a common scale would not change the selections, but the intended normalization should be stated or the values corrected.","section":"Eq. (5) and Section IV-A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Several references cited in the introductory survey of devices collecting private sensor data (wireless power transfer [15] and [16], tandem mass spectra [17], and GUI-agent papers [18]–[20]) appear unrelated to the text they are attached to; please re-check the citation placement for these entries.","section":"References [15]–[20]","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"On ActionSense and WEAR the test windows are the last 20% of each subject's windows, so test subjects also appear in training; the text discloses this, but one sentence noting that absolute accuracies under this protocol are not directly comparable with subject-disjoint evaluation would help readers interpret the numbers.","section":"Section IV-A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The selection rows of Table I use K = 20 on UCI-HAR and K = 3 and K = 6 on ActionSense and WEAR, while the main TIDE row uses K = 20 throughout (so TIDE's selection is effectively full participation on ActionSense and WEAR); this distinction is easy to misread and should be restated at the point where the 'Score-based selection, same K' rows are discussed.","section":"Table I caption and Section IV-B","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'RELIEF is the most accurate at 5%' is ambiguous about whether 5% refers to the gyroscope share or to another quantity; please clarify what is being varied in that run.","section":"Section IV-B, Fig. 5(b)","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The short version: TIDE is a legitimate engineering paper — the central mechanism, merging each sensor's encoder only within the pod of devices that hold it, demonstrably works — but the headline claim of being \"most accurate at half its own communication on all three datasets\" is softer than it reads. Two of the three margins sit at or inside the seed variation the paper itself reports, and the comparison field includes a handicapped baseline.\n\nWhat's new: per-sensor encoder plus auxiliary head, averaged only within the sensor's pod, with shared fusion and task head, under periodic availability with rarity-aware selection. The individual pieces predate it — RELIEF aggregates modality blocks within holder sets and MA-CLAMP averages layers over training clients, and the authors credit both — but the combination is new, and the ablation shows the gain is real. Intra-pod merging alone adds 7.58 points on ActionSense and nearly halves the worst-case drop on two datasets. The sensor-scarcity experiment is the cleanest evidence: at 10% gyroscope coverage, TIDE reaches 51.58% gyro-only accuracy against 16.42% for naive averaging.\n\nWhere it's soft: the equal-communication comparison. Budget B is TIDE's own communication, a defensible choice, and B/2 is reported as well. But the \"most accurate at B/2 on all three\" sentence carries more weight than the numbers support. On UCI-HAR the 1.07-point margin over RELIEF sits inside the reported seed variation; on WEAR the 1.49-point margin over FedMVD is borderline; only ActionSense is decisive. And on the two small-margin datasets the field includes PFHAR, reproduced from random initialization without its large-scale pretraining — an acknowledged modification, never bounded. PFHAR without pretraining still posts 89.87 on UCI-HAR, within 1.30 of TIDE's 91.17, so a pretrained copy plausibly reorders the B/2 results. The approximate round-to-GiB conversion for staged Harmony and FedHKD is a genuine blemish (FedHKD shows 3.46% at B on WEAR versus 72.98% at termination), though the leading rivals at B are non-staged and unaffected. No code is released; only independent reproduction settles the baseline question.\n\nNet: mechanism real, writing honest, limitations disclosed. The abstract and conclusion should bound the B/2 superlative to the setting that supports it. Worth a serious referee — a solid paper in a practical subfield that would benefit from one external reproduction of the baseline numbers.","headline":"TIDE's intra-pod merging mechanism is real and the ablation is honest, but the 'most accurate at B/2 on all three datasets' headline leans on margins that are within seed variation on two of three datasets and a PFHAR baseline reproduced without its defining pretraining.","tokens_in":16710,"tokens_out":4793,"duration_ms":31960},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Averaging each sensor's model only across devices that hold it — the 'pod' rule — makes the federation the most accurate on three activity datasets at half its own communication cost.","keywords":["federated learning","multimodal sensing","Internet of Things","device heterogeneity","client scheduling","human activity recognition","intra-pod aggregation","periodic availability"],"falsifier":"Rerun the comparison of Table I with PFHAR warm-started from its published pretrained initialization, GraMFedDHAR at its original noise multiplier, and Harmony and FedHKD communication counted in actual transmitted bytes; if TIDE's equal-budget margins on UCI-HAR and WEAR fall within seed variation, the headline accuracy claim fails.","tokens_in":17195,"feed_emoji":"⌚","tokens_out":15558,"duration_ms":89085,"temperature":0.7,"pith_summary":"This paper claims that the failure of multimodal federated learning on heterogeneous IoT devices is mostly an aggregation problem: when devices carry different sensors and appear periodically, global averaging lets a device without a sensor overwrite that sensor's learned parameters, and a single-sensor device never learns how its sensor relates to the others. TIDE fixes the aggregation rule: each sensor has its own encoder and auxiliary classification head, and those parameters are averaged only over the 'pod' of devices that actually hold the sensor, while the fusion and task layers stay shared. On three human-activity-recognition datasets against eight baselines, the paper reports that this scheme is the most accurate at half its own communication cost on all three, leads at equal communication on ActionSense, and is ahead within seed variation on the other two. A sympathetic reader should care because the fix is a change to how parameters are merged, not a new local-training recipe, so it transfers to any sensor-partitioned federation.","feed_headline":"Update each sensor's model only among devices that hold it","feed_subtitle":"The pod-wise scheme is the most accurate on three activity datasets at half its own communication cost.","key_machinery":"The pod $C_m$ — the set of devices that hold sensor $m$ — together with the decoupled merging rule of Eq. (11), which averages the sensor's encoder and auxiliary head only over the selected members of the pod and otherwise carries the previous value forward. The shared fusion layer and task head, averaged over all selected devices, provide the federation-wide part; per-sensor auxiliary heads make each encoder discriminative on its own so that the pod has something worth averaging; server-held class prototypes, a cross-sensor alignment loss, modality dropout, and a rarity-aware selection score round out the scheme. The equal-cost objective Acc@B — accuracy at a fixed communication budget — is what lets the paper compare methods at half or equal communication rather than at their own stopping points.","core_discovery":"TIDE's central claim is that $\\theta_m$, the encoder together with the auxiliary head of each sensor $m$, should be updated only by devices in its pod $C_m$, weighted by data size, and simply kept at its previous value in rounds where no pod member is selected (Eq. 11); the fusion layer and the task head are averaged over all selected devices. This removes the overwriting that plain FedAvg causes — no selected device outside $C_m$ contributes a copy of $\\theta_m$ that never saw real data for sensor $m$ — and, because multi-sensor devices belong to several pods, those devices end up training the encoders that their single-sensor neighbours rely on. The paper reports that on ActionSense this reaches 78.06%, 7.63 points above the next method; that at half TIDE's own communication it is the most accurate on all three datasets; that at equal budget it leads on ActionSense and is ahead within seed variation on UCI-HAR and WEAR; and that the ablation credits most of the gain to the auxiliary heads plus intra-pod merging, which together roughly halve the worst-case accuracy drop when a sensor is missing on two datasets.","pith_inferences":["Because the selection score matters only on UCI-HAR and trails random selection at small K on WEAR, the selection module looks incidental to the method's contribution; the pod-wise aggregation rule is likely the portable part, and one could test it with uniform random selection everywhere without expecting much change.","The pod idea is not tied to sensors: the same 'merge a component's parameters only over the subpopulation that actually observes it' rule could apply to federated learning with heterogeneous data views, such as languages, camera views, or input modalities, and that extension is directly testable.","The reported margins are sensitive to how baselines are reproduced, so a fair re-run with exact byte counts and pretrained initializations would either confirm the equal-budget advantage or shrink it; the paper's own footnote about the approximate GiB conversion for staged methods flags this directly.","A natural stress test the paper does not run is scaling to federations with many rare sensors and very small pods, where intra-pod averaging has almost no partners and the rule approaches keeping local copies; the 10%-gyroscope experiment is a first step in that direction but varies only one sensor's prevalence."],"forward_implications":["Any multimodal federation with missing sensors can adopt pod-wise merging without changing local training, and the ablation shows it accounts for most of the gain: +0.80, +7.58, and +1.14 points on UCI-HAR, ActionSense, and WEAR on top of all alignment losses.","Single-sensor devices inherit working encoders: with a gyroscope on only 10% of devices, gyroscope-only accuracy under TIDE is 51.58% versus 16.42% under naive averaging.","Equal-cost evaluation reverses the UCI-HAR ranking: RELIEF's final-accuracy lead of 1.42 points costs 3.7 times TIDE's communication, and at half TIDE's budget TIDE is the most accurate on all three datasets.","Worst-case sensor-drop nearly halves on two datasets (32.47 to 16.14 points on UCI-HAR, 39.45 to 15.47 on ActionSense) when the auxiliary heads and intra-pod merging are added.","The paper's ablation identifies the auxiliary heads and intra-pod merging, not the prototypes, cross-sensor loss, or participation term, as the source of the gain; the other losses move accuracy within seed variation."],"supporting_citations":[{"why":"Closest predecessor: it already aggregates the fusion layer's modality blocks within the devices holding each modality; it is the main rival that finishes above TIDE at termination on UCI-HAR while spending 3.7 times the communication.","marker":"[34]"},{"why":"Staged heterogeneous-multimodal baseline whose rounds are converted to GiB only approximately, so it anchors the equal-communication comparison.","marker":"[26]"},{"why":"Staged hierarchical-disentanglement baseline whose communication is also approximated in GiB and which stays near chance on ActionSense in the authors' reproduction.","marker":"[27]"},{"why":"Multi-view domain-alignment baseline that is the closest competitor on WEAR at termination (79.85% versus 80.42%).","marker":"[28]"},{"why":"Per-sensor classifier baseline with the best worst-case sensor-drop; TIDE claims the second-smallest drop behind it.","marker":"[29]"},{"why":"Foundation-model baseline reproduced from random initialization without its large-scale pretraining; the fairness of that reproduction is load-bearing for the reported margins.","marker":"[32]"},{"why":"Differentially private baseline run with an adjusted noise multiplier (0.5 on WEAR, where 1.0 diverged) and not converged on UCI-HAR or ActionSense.","marker":"[33]"},{"why":"UCI-HAR dataset with 100 clients in three sensor tiers; the equal-budget claim depends on this comparison, where RELIEF's final accuracy leads.","marker":"[37]"},{"why":"ActionSense kitchen-task dataset where TIDE posts its largest margin, 78.06% against the next method's 70.43%.","marker":"[38]"},{"why":"WEAR outdoor inertial-plus-video dataset with three tiers, where TIDE leads at termination but trails random selection at small K.","marker":"[39]"}],"fun_headline_variants":["Federated learning that updates each sensor only among devices that hold it","Pod-wise sensor encoders beat eight baselines on three IoT datasets","TIDE keeps sensor models inside their pods, lifting activity recognition","Half the communication, top accuracy on three multimodal IoT datasets","Sensor-aware federated learning: no overwrites, better activity recognition"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that the baseline reproductions are not handicapped: if PFHAR run from its pretrained weights, GraMFedDHAR at its original noise setting, or Harmony and FedHKD with exact byte counts would close the reported gaps, the equal-communication margins would shrink.","fun_headline_variants_meta":{"raw":{"variants":["Federated learning that updates each sensor only among devices that hold it","Pod-wise sensor encoders beat eight baselines on three IoT datasets","TIDE keeps sensor models inside their pods, lifting activity recognition","Half the communication, top accuracy on three multimodal IoT datasets","Sensor-aware federated learning: no overwrites, better activity recognition"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":5.4e-05,"raw_usage":{"cost":5.4054e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":5.4054e-05,"upstream_inference_prompt_cost":3.185e-05,"upstream_inference_completions_cost":2.2204e-05},"total_tokens":1188,"prompt_tokens":1066,"completion_tokens":122,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":895,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":46}},"tokens_in":1066,"tokens_out":122,"cost_source":"openrouter_usage","duration_ms":5444,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":46,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-03T15:50:30.627508+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Rerun the comparison of Table I with PFHAR warm-started from its published pretrained initialization, GraMFedDHAR at its original noise multiplier, and Harmony and FedHKD communication counted in actual transmitted bytes; if TIDE's equal-budget margins on UCI-HAR and WEAR fall within seed variation, the headline accuracy claim fails.","supporting_citations":[{"cited_title":"RELIEF: Turning missing modalities into training acceleration for federated learning on heterogeneous IoT edge,","cited_arxiv_id":null,"evidence_quote":"Closest predecessor: it already aggregates the fusion layer's modality blocks within the devices holding each modality; it is the main rival that finishes above TIDE at termination on UCI-HAR while spending 3.7 times the communication."},{"cited_title":"Harmony: Heterogeneous multi-modal federated learning through disentangled model training,","cited_arxiv_id":null,"evidence_quote":"Staged heterogeneous-multimodal baseline whose rounds are converted to GiB only approximately, so it anchors the equal-communication comparison."},{"cited_title":"Towards efficient heterogeneous multi-modal federated learning with hierarchical knowledge disentanglement,","cited_arxiv_id":null,"evidence_quote":"Staged hierarchical-disentanglement baseline whose communication is also approximated in GiB and which stays near chance on ActionSense in the authors' reproduction."},{"cited_title":"Multimodal fusion using multi-view domains for data heterogeneity in federated learning,","cited_arxiv_id":null,"evidence_quote":"Multi-view domain-alignment baseline that is the closest competitor on WEAR at termination (79.85% versus 80.42%)."},{"cited_title":"Cross-modal federated human activity recognition,","cited_arxiv_id":null,"evidence_quote":"Per-sensor classifier baseline with the best worst-case sensor-drop; TIDE claims the second-smallest drop behind it."},{"cited_title":"PFHAR: Practically adopting multi-modal foundation model for human activity recognition through edge-cloud collaborative learning,","cited_arxiv_id":null,"evidence_quote":"Foundation-model baseline reproduced from random initialization without its large-scale pretraining; the fairness of that reproduction is load-bearing for the reported margins."},{"cited_title":"GraMFedDHAR: Graph based mul- timodal differentially private federated human activity recognition,","cited_arxiv_id":null,"evidence_quote":"Differentially private baseline run with an adjusted noise multiplier (0.5 on WEAR, where 1.0 diverged) and not converged on UCI-HAR or ActionSense."},{"cited_title":"A public domain dataset for human activity recognition using smartphones,","cited_arxiv_id":null,"evidence_quote":"UCI-HAR dataset with 100 clients in three sensor tiers; the equal-budget claim depends on this comparison, where RELIEF's final accuracy leads."},{"cited_title":"ActionSense: A multimodal dataset and recording framework for human activities using wearable sensors in a kitchen environment,","cited_arxiv_id":null,"evidence_quote":"ActionSense kitchen-task dataset where TIDE posts its largest margin, 78.06% against the next method's 70.43%."},{"cited_title":"WEAR: An outdoor sports dataset for wearable and egocentric activity recognition,","cited_arxiv_id":null,"evidence_quote":"WEAR outdoor inertial-plus-video dataset with three tiers, where TIDE leads at termination but trails random selection at small K."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}