When the Cross-Silo Federation Goes Offline: Continual Learning for Site Onboarding with Limited Unlabeled Data

1 Introduction

• Late-joining sites can arrive after established participants go offline with unlabeled data mixing known and unseen conditions, while class-incremental training risks catastrophic forgetting.Replay, regularization, and distillation are constrained because old records remain with their owners, estimates must be prepared before disconnection, or new-site data may differ in distribution.Source: §1 · p. 2
• The joining site autonomously screens novelty, groups flagged records, and expands the classifier in one shared representation, without requiring established sites to reconnect.The same representation supports novelty scoring, clustering, and means describing old classes.Source: §1 · p. 2
• The procedure is evaluated on real ultrasonic-welding process data across class-arrival patterns against distillation- and regularization-based continual-learning methods.The evaluation concerns unlabeled, few-record onboarding while established sites remain offline.Source: §1 · p. 2
• Federation-level prototypes and joining-site records differ in accuracy by at most 0.022, while both keep old- and new-class accuracy within 0.026 where measured alternatives trade one against the other.The two sources are presented as a privacy-indexed choice rather than an ablation ladder.Source: §1 · p. 3

2 Related work

• Federated continual-learning protocols commonly assume labeled incremental tasks, continuing client participation, and a proxy server, exemplar memory, or server-side generator, unlike the offline late-site setting.The paper therefore compares against centralized continual-learning methods adapted to this setting.Source: §2 · p. 3
• Open-set recognition rejects inputs outside known labels, while reconstruction models suit cases where future-class examples are unavailable during training; prior federated disagreement methods assume concurrent clients.The paper trains one federated expert per known class, and its screen also selects records anchoring old classes.Source: §2 · p. 3
• Federated continual novel-class learning discovers classes across participating clients and continues training to preserve earlier classes, but assumes sites remain online and have enough data for cross-client agreement.The joining site here must decide alone from a small pool after established sites go offline.Source: §2 · p. 3
• Few-shot class-incremental learning can use prototypes with a frozen representation, but its new classes are already named and labeled.Federated generalized category discovery instead treats clustering unlabeled data into known and novel categories as the end of the task.Source: §2 · p. 3
• Cross-silo settings involve a small number of persistently identified organizations that cannot release their data, with scarce local labels and disclosure-sensitive records.The paper instantiates its protocol on ultrasonic welding, a previously studied centralized setting that does not impose offline-participant or disclosure constraints.Source: §2 · p. 3

3.1 Setting

• The base federation has M sites with private datasets over K old classes and trains a model without moving raw records.The setting defines Yold as the shared class set.Source: §3.1 · p. 4
• After deployment, site M+1 joins with an unlabeled pool mixing old classes and a disjoint, initially unknown set of new classes.The size and identity of the new-class set are unknown on arrival.Source: §3.1 · p. 4
• The joining site expands the classifier to K + R outputs using its available data and the summaries retained by the federation.Adaptation takes place at the joining site alone; established sites need not reconnect and their records are not transferred.Source: §3.1 · p. 4
• The federation retains the shared model, K reconstruction experts, detection constants, and one representation-level prototype per site and old class.When permitted, the joining site also keeps up to b of its own records per old class.Source: §3.1 · p. 4

3.2 Evaluation protocol

• Evaluation uses a common held-out test set and reports old-class accuracy before adaptation, old-class accuracy afterward, and accuracy on arriving conditions.Aold is averaged over the established sites; Anew measures performance on the new conditions.Source: §3.2 · p. 4
• Forgetting is Fold = A0 old − Aold, the old-class accuracy drop on the established-site population; the joining site's old-class accuracy is reported separately.The joining site has no before-adaptation figure for its own old classes.Source: §3.2 · p. 4

4.1 The shared representation

• The early hidden-layer representation supports screening, grouping, and old-class replay because it separates known classes, keeps same-condition records nearby, and lets sites summarize classes without sending records.The representation is chosen to be deep enough not to carry input variables themselves but shallow enough to encode shared process structure rather than only the base classes.Source: §4.1 · p. 5

4.2 Screening: identifying unfamiliar records

• Class-specific denoising autoencoders score records by whether a specialist reconstructs them well, and the joining site flags records whose normalized score exceeds a threshold based on known-class scores.The score is calibrated against each expert’s typical class-specific reconstruction error because classes differ in reconstruction difficulty.Source: §4.2 · p. 5

4.3 Grouping: estimating how many conditions have arrived

• The joining site reduces flagged representations with PCA and clusters them using DBSCAN, which requires no target cluster count and leaves low-density points unassigned.Each recovered group opens one new output, without consulting labels.Source: §4.3 · p. 5

4.4 Expansion and the evidence it requires

• Expansion adds one output per recovered group while freezing the old-output weights, but retaining old-class accuracy still requires old-class evidence in the adaptation objective.That evidence can be federation-level class prototypes or unlabeled records retained locally by the joining site; prototypes are class means in representation space, while local records are selected through the novelty screen.Source: §4.4 · pp. 5–6

4.5 A menu indexed by disclosure

• The adaptation configurations are a menu of disclosure choices rather than successive steps toward one best setting.The configurations hold different components fixed and require different disclosures.Source: §4.5 · p. 6, Table 1 · p. 6

• The dataset contains 270 ultrasonic-welding cycles with 252 engineered features across nine balanced classes; six new- and worn-tool classes are old, while three damaged-tool classes arrive later.The arriving classes pair damaged tools with familiar surface states, differing from known conditions along one factor.Source: §5 · p. 6
• Each class’s 30 records are split into six common test records, six per established site for old classes, and joining-site data; all accuracies use the 54-record test set.The joining site receives 36 familiar and 72 unfamiliar records.Source: §5 · p. 6
• The base classifier is a multilayer perceptron trained for 20 federated rounds, and adaptation uses 20 local rounds, prototype-loss weight one, and three retained records per old class.The shared representation has 150 dimensions, with one denoising autoencoder per old class and detection quantile q = 0.95.Source: §5 · pp. 6–7
• LWF and EWC are compared under the same architecture, split, schedule, and arrival patterns, each at its most favorable balanced weight from a sweep.The paper’s configurations use a fixed prototype weight of one and are not swept.Source: §5 · p. 7
• The evaluation includes verified-label arrivals across all seven nonempty arrival patterns and a separate end-to-end run with no labels consulted; no jointly trained nine-class reference is available.Old-class accuracy differences below 0.028 are not treated as separating configurations because one test record is worth 0.028.Source: §5 · p. 7

6.1 Screening

• 0.975±0.027 recall, 0.750±0.068 acceptance, and 0.941± 0.018 ROC-AUC are achieved by the novelty detector at the federated threshold on the joining-site pool.On average it misses under two unfamiliar records and flags one familiar record in four in error.Source: §6.1 · p. 7

6.2 The evidence that expansion requires

• 0.931±0.035 old-class accuracy with forgetting of 0.008±0.020 is reached by DUAL in controlled arrivals; GLOBAL and LOCAL reach 0.909±0.048 and 0.924±0.042, respectively.The replay configurations’ largest old-class accuracy separation is 0.022, below the 0.028 contribution of one test record.Source: §6.2 · p. 7

6.3 The procedure end to end

• 0.889 ± 0.055 old-class accuracy and 0.889±0.039 new-class accuracy are reached by DUAL in the full autonomous procedure, compared with 0.149±0.084 new-class accuracy for fine-tuning.Across replay configurations, old-class accuracy ranges from 0.868±0.036 to 0.889±0.055 despite errors propagating through the pipeline.Source: Table 2 · p. 8, §6.3 · p. 8

6.2 The evidence that expansion requires

• DUAL reaches 0.931±0.035 old-class accuracy and 0.905±0.041 new-class accuracy, while balanced LWF and EWC have lower old-class accuracy and higher new-class accuracy.LWF reaches 0.707±0.039 old-class and 0.956±0.023 new-class accuracy; EWC reaches 0.809 ± 0.054 old-class and 0.935 ± 0.042 new-class accuracy.Source: §6.2 · pp. 7–8
• Replay keeps old- and new-class accuracy within 0.026, whereas every other table entry separates them by 0.126 to 0.419; LOCAL retains this balance with six labeled records per arriving condition and three retained per old class.Source: §6.2 · p. 8

• The reported margins do not support interpreting Table 1’s configurations as successive ablation steps.Source: §7 · p. 8
• GLOBAL requires established sites to release class means, LOCAL requires the joining site to retain records, and DUAL applies when both disclosures are permitted.Either GLOBAL or LOCAL can be chosen when the other party’s disclosure is foreclosed, with comparable accuracy reported for the alternatives.Source: §7 · p. 9

8 Limitations and future work

• The protocol requires sites to measure the same quantities consistently; re-instrumentation or a partly different feature set breaks correspondence and prevents admission as it stands.Aligning heterogeneous feature spaces is identified as a future extension.Source: §8 · p. 9
• Rare arriving conditions and sites that differ by machine or operator are untested pressures that could shift screening and the representation summarized by prototypes.Every class contributes the same number of records, and every site observes one process on one machine.Source: §8 · p. 9
• The study assumes one shared classification task and one expansion per joining site; distinct related tasks and process drift across arrivals lie outside its measurements.Reported effect sizes come from one industrial instantiation, and the reference methods are compared as published rather than with the replay configurations’ added supports.Source: §8 · p. 9

9 Conclusion

• The joining site absorbs unlabeled, previously unseen conditions in one shared representation, with no labels consulted and no established site online.Old-class evidence from two dissimilar sources supports comparable accuracy, allowing the choice to reflect permitted disclosure rather than performance.Source: §9 · p. 9

A Screening

• 0.941±0.018 ROC-AUC for the reconstruction bank exceeded 0.884±0.033 for maximum softmax probability and 0.875±0.040 for the energy score.The bank ranked ahead of the energy score on 5 of 5 splits and maximum softmax probability on 4 of 5, and its per-class expert errors explain why records were flagged.Source: A Screening · p. 12, Table A1 · p. 12

B Reference-method sweeps

• Figure A1 reports 12 LWF weights spanning 0.05 to 200 and 6 EWC weights spanning 108 to 1014, each averaged over 5 seeds.The table uses weights 0.25 for LWF and 1011 for EWC.Source: B Reference-method sweeps · p. 12
• Both reference methods were extended until the harmonic mean of old- and new-class accuracies peaked and began to decline; replay configurations used a fixed prototype weight and were not swept.The reference methods were the only table entries selected by this sweep criterion.Source: B Reference-method sweeps · p. 12
• Replay configurations sit above both LWF and EWC sweeps at every weight, retaining substantially more old-class accuracy at comparable new-class accuracy.No setting of either reference-method weight reaches the region occupied by the replay configurations.Source: B Reference-method sweeps · p. 12

C How much data the joining site needs

• New-class accuracy is 0.556±0.104 with one labeled record per condition, 0.800±0.115 with three, and 0.900±0.046 with six, while remaining nearly flat in b.The grid varies n from 1 to 6 and b from 0 to 6, with all three conditions arriving at once and results averaged over five splits.Source: C How much data the joining site needs · p. 12
• Old-class accuracy is 0.513±0.209 at b=0 and falls to 0.253±0.083 at n=6; it rises to 0.876±0.053 at b=1 and 0.910±0.060 at b=2.The b=0 row is the only region where the model loses old classes.Source: C How much data the joining site needs · p. 13
• With n ≥3 and b ≥1, every LOCAL budget cell is within 0.054 of the best cell (n=6, b=3: old 0.917±0.065; new 0.900±0.046).The paper identifies one or two retained records per old class, together with three examples of each new condition, as enough to admit the site.Source: C How much data the joining site needs · pp. 13–14

D Per-split grouping outcomes

• Across five complete runs, Table A2 reports recovered group counts, partition agreement, and flagged records left unassigned by DBSCAN; unassigned records are excluded from new-class training.ARI is computed between the recovered partition and reference conditions on assigned records.Source: D Per-split grouping outcomes · p. 14, Table A2 · p. 14

E Reproducibility and compute

• Reported variation is the sample standard deviation across independently seeded runs and is descriptive; the authors draw no inferential conclusion from it.Source: E Reproducibility and compute · p. 14

F Notation

• The notation defines M and K as established-site and old-class counts, R as added conditions, U as the joining site's unlabeled pool, and Fold as A0 old − Aold.Source: F Notation · p. 15