Federated Concept-Based Models: Interpretable models with distributed supervision
Authors: Dario Fenoglio, Arianna Casanova, Francesco De Santis, Gabriele Dominici, Johannes Schneider, Pietro Barbiero, Giovanni De Felice, Marc Langheinrich, Martin Gjoreski
Organizations: Universit`a della Svizzera italiana Lugano, Switzerland · University of Liechtenstein Vaduz, Liechtenstein · Politecnico di Torino Turin, Italy · IBM Research Zurich, Switzerland
Concept-based Models (CMs) enhance interpretability in deep learning by grounding predictions in human-understandable concepts. However, concept annotations are costly and rarely available at scale within a single data source. Federated Learning (FL) could alleviate this limitation by enabling cross-institutional training over concept annotations distributed across multiple data owners. Yet, FL lacks interpretable modeling paradigms. Integrating CMs with FL is non-trivial: although FL supports heterogeneous and non-stationary client participation, it typically assumes a fixed shared architecture, whereas CMs may require architectural adaptation as the available concept set evolves. We propose Federated Concept-based Models (F-CMs), a new methodology for deploying CMs in evolving FL settings. F-CMs aggregate concept-level information across institutions and efficiently adapt the model architecture to changes in concept supervision while preserving privacy. Empirically, F-CMs maintain accuracy and intervention effectiveness comparable to training settings with full concept supervision, while outperforming on average non-adaptive federated baselines. Notably, F-CMs enable interpretable inference on concepts unavailable to a given institution, a key novelty over existing approaches.
Figures & tables
Figure 1: F-CMs overview. At each round t , F-CMs perform three steps: (1) Graph aggregation: the server combines client graphs into a shared graph Gt ; (2) Architecture adaptation: the shared CM adapts to Gt and/or newly observed concepts; (3) Training: clients update only supervised concept/task modules (dark grey), leaving others frozen (light grey), and the server aggregates the corresponding updates. The figure shows two rounds, where new clients introduce new concepts.
Setting
Method
Asia
Sachs
Alarm
Insurance
Hailfinder
SIIM-Pn.
CheXpert
T
C. Cov.
T
C. Cov.
T
C. Cov.
T
C. Cov.
T
C. Cov.
T
C. Cov.
T
C. Cov.
Cent. (Upper Bound)
OpaqNN
80.7 ± 1.0
100
77.7 ± 1.0
100
74.0 ± 0.3
100
76.3 ± 1.1
100
73.0 ± 0.8
100
74.1 ± 0.6
100
72.0 ± 0.3
100
CBM
80.0 ± 1.0
100
77.3 ± 0.9
100
73.9 ± 0.5
100
76.3 ± 1.4
100
74.2 ± 0.7
100
73.9 ± 0.3
100
70.7 ± 0.5
100
CEM
80.4 ± 1.1
100
77.3 ± 0.8
100
71.6 ± 0.4
100
76.0 ± 1.2
100
73.3 ± 0.6
100
73.1 ± 0.7
100
66.6 ± 0.7
100
CGM
79.8 ± 1.2
100
76.2 ± 1.2
100
71.5 ± 0.9
100
73.4 ± 1.1
100
72.2 ± 0.8
100
68.8 ± 0.8
100
71.6 ± 0.3
100
C 2 BM
80.3 ± 1.0
100
77.9 ± 1.0
100
72.8 ± 0.5
100
74.8 ± 0.9
100
73.5 ± 0.6
100
72.8 ± 0.1
100
69.4 ± 0.7
100
Table 1: Task accuracy (%) and concept coverage (%) with respect to the ground-truth concepts causally relevant for the task, i.e., concepts with a directed path to the task in the ground-truth graph or the available proxy. Methods that improve the average performance are shown in bold. Results are reported as mean ± standard error over five runs.
Figure 2: Label accuracy (%) on downstream variables (including the task) following interventions on concepts at deeper levels of the ground-truth graph hierarchy. Dashed lines denote interventions that are not possible for a given modality due to the absence of predictions for the concept.
Figure 3: Convergence under temporal non-stationarity. Average task loss across rounds for C 2 BM and CEM, where new clients introduce unseen concepts. F-CMs (ours) adapt the architecture and converge faster and to lower loss than S-F-CMs (no adaptation) and S-F-CMs Reinit (retraining). Curves are truncated at the best loss-round. The legend reports the fraction of modified parameters.
Appendix figures & tables35 assets
Supplementary material from the paper’s appendix.
Appendix
Bayesian network
Nodes
Edges
Domain
Asia
8
8
Medical diagnosis
Sachs
11
17
Protein signaling
Alarm
37
46
Medical monitoring
Insurance
27
52
Insurance risk
Hailfinder
56
66
Weather prediction
Appendix
Table 2: Summary of bnlearn Bayesian networks used in experiments.
Setting
Model
Asia
Sachs
Alarm
Insurance
Hailfinder
SIIM Pn.
CheXpert
Cent. (Upper Bound)
OpaqNN
91.6 ± 0.4
75.3 ± 0.5
91.3 ± 0.1
79.5 ± 0.2
61.4 ± 0.3
74.1 ± 0.2
89.4 ± 0.2
CBM
91.5 ± 0.4
74.8 ± 0.7
91.0 ± 0.1
79.7 ± 0.2
69.7 ± 0.1
74.1 ± 0.1
89.1 ± 0.3
CEM
91.3 ± 0.4
73.9 ± 0.5
89.9 ± 0.3
77.1 ± 0.3
66.3 ± 0.3
72.8 ± 0.1
89.0 ± 0.1
CGM
91.0 ± 0.5
72.3 ± 0.5
89.5 ± 0.2
72.7 ± 0.3
63.8 ± 0.3
66.8 ± 0.8
89.2 ± 0.1
C 2 BM
91.3 ± 0.4
74.4 ± 0.5
90.1 ± 0.4
77.5 ± 0.2
65.4 ± 0.6
72.8 ± 0.2
89.1 ± 0.1
Loc.
OpaqNN
87.0 ± 3.4
67.9 ± 0.5
85.3 ± 1.3
73.3 ± 1.2
53.0 ± 1.8
67.2 ± 0.6
87.5 ± 1.4
Appendix
Table 3: Concept Accuracy.
Figure 4: Task accuracy (%) following interventions on concepts at increasing depths of the ground-truth graph hierarchy. Dashed lines denote interventions that are not possible for a given training setting due to the absence of predictions for the corresponding concept.
Figure 5: Label accuracy (%) on downstream variables (including the task) following interventions on concepts at increasing depths of the ground-truth graph hierarchy. Dashed lines denote interventions that are not possible for a given training setting due to the absence of predictions for the corresponding concept.
Figure 6: Task accuracy (%) following interventions on concepts at increasing depths of the ground-truth graph hierarchy. Dashed lines denote interventions that are not possible for a given training setting due to the absence of predictions for the corresponding concept.
Figure 7: Label accuracy (%) on downstream variables (including the task) following interventions on concepts at increasing depths of the ground-truth graph hierarchy. Dashed lines denote interventions that are not possible for a given training setting due to the absence of predictions for the corresponding concept.
Figure 8: Task accuracy (%) following interventions on concepts at increasing depths of the ground-truth graph hierarchy. Dashed lines denote interventions that are not possible for a given training setting due to the absence of predictions for the corresponding concept.
Figure 9: Label accuracy (%) following interventions on concepts at increasing depths of the ground-truth graph hierarchy across CMs instantiations.
Figure 10: Task accuracy (%) following interventions on concepts at increasing depths of the ground-truth graph hierarchy across CMs instantiations.
Figure 11: Label and task accuracy (%) on CheXpert following interventions on concepts at increasing depths of the ground-truth graph hierarchy across CMs instantiations.
Figure 12: Convergence under temporal non-stationarity for additional model instantiations. Average validation task loss across rounds for CBM and CGM , where new clients introduce previously unseen concepts. F-CMs (ours) adapts the architecture modularly and converges faster and to lower loss than S-F-CMs (no adaptation) and S-F-CMs Reinit (full retraining). Curves are truncated at the best loss-round; the omitted tail corresponds to no-improvement rounds. The legend reports the fraction of modified parameters.
Model
Asia
Sachs
Alarm
Insurance
Hailfinder
SIIM Pn.
OpaqNN
18.28 ± 2.56
38.32 ± 6.40
42.76 ± 5.82
28.90 ± 4.24
17.37 ± 2.93
5.46 ± 0.67
CBM
62.55 ± 12.93
70.00 ± 15.48
48.44 ± 7.88
26.90 ± 3.58
19.24 ± 4.21
9.27 ± 1.26
CEM
61.30 ± 11.83
81.11 ± 18.54
57.30 ± 8.84
30.44 ± 4.48
17.80 ± 3.13
21.61 ± 3.01
CGM
32.19 ± 3.17
43.09 ± 7.05
45.87 ± 10.04
19.68 ± 3.39
8.37 ± 1.45
10.29 ± 1.33
C 2 BM
28.56 ± 4.54
44.82 ± 8.49
31.06 ± 4.62
19.24 ± 2.78
12.73 ± 2.05
8.49 ± 1.13
Appendix
Table 4: Fraction of parameters changed (%) due to temporal non-stationarity .
Data
Model
d_inst.
d_par.
agg.
n_rounds
tot.
Asia
F-CMs (CEM)
0.227 ± 0.001
0.002 ± 0.000
–
42.583 ± 2.256
229.726 ± 13.609
SIIM-Pn.
F-CMs (CEM)
0.691 ± 0.093
0.003 ± 0.000
–
35.350 ± 5.188
506.230 ± 75.622
Asia
F-CMs (C 2 BM)
0.244 ± 0.001
0.003 ± 0.000
0.063 ± 0.051
61.500 ± 3.679
370.800 ± 24.924
SIIM-Pn.
F-CMs (C 2 BM)
0.672 ± 0.102
0.004 ± 0.000
0.037 ± 0.003
56.550 ± 7.278
922.543 ± 120.707
Appendix
Table 5: Average runtime (seconds) of the additional operations introduced by F-CMs , reported as mean ± 95% confidence interval over three runs. Here, d_inst. denotes the time spent instantiating the architecture after client drift, d_par. the time spent initializing the post-drift architecture with previously learned parameters, agg. the time spent aggregating client graphs, n_rounds the number of training rounds, and tot. the total training time. The results show that graph aggregation and modular adaptation contribute only a very small fraction of the total runtime.
New params./concept
Added params.
Memory (MiB)
Forward MAC increase
Dataset
Concepts added
CEM
C 2 BM
CEM
C 2 BM
CEM
C 2 BM
CEM
C 2 BM
Asia
3
1,346
610
5,130
1,982
0.066
0.044
42.0%
20.7%
Alarm
8
3,843
1,763
48,513
15,516
0.745
0.335
33.1%
21.4%
Appendix
Table 6: Scaling under concept accumulation. Growth of the concept-dependent component for CEM and C 2 BM as new concepts are incorporated. “New params.” denotes the parameters of a newly instantiated concept module, while “Added params.” also includes affected interfaces. Memory is the post-drift FP32 parameter footprint. The fixed input encoder is excluded from all measurements.
Figure 13: Interface-initialization ablation on Asia . Average validation task loss across federated rounds for C 2 BM (left) and CGM (right), under concept-space expansions from 10 – 50% to 100% structural concept coverage (top) and from 51 – 75% to 100% coverage (bottom). We compare the current initialization strategy with partial parameter reuse, using either random or zero initialization for newly introduced parameters. The vertical dashed line marks the structural update. Curves and shaded regions report mean ± standard deviation over three seeds.
Figure 14: Interface-initialization ablation on Alarm . Average validation task loss across federated rounds for C 2 BM (left) and CGM (right), under concept-space expansions from 10 – 50% to 100% structural concept coverage (top) and from 51 – 75% to 100% coverage (bottom). We compare the current initialization strategy with partial parameter reuse, using either random or zero initialization for newly introduced parameters. The vertical dashed line marks the structural update. Curves and shaded regions report mean ± standard deviation over three seeds.
Privacy Budget
ε=10
ε=5
ε=1
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
OpaqNN
80.03 ± 0.68
90.37 ± 0.70
79.19 ± 0.97
89.43 ± 1.32
77.31 ± 0.96
89.15 ± 1.07
C 2 BM
79.37 ± 1.01
86.89 ± 2.78
79.20 ± 1.16
86.89 ± 2.63
76.22 ± 2.24
85.81 ± 2.06
CBM
79.11 ± 1.79
89.32 ± 1.46
79.03 ± 1.87
89.13 ± 1.49
72.73 ± 6.52
88.07 ± 2.16
CEM
80.19 ± 1.08
86.26 ± 2.72
80.17 ± 1.63
87.26 ± 2.74
78.12 ± 2.68
85.67 ± 2.47
CGM
78.67 ± 2.39
88.59 ± 1.34
77.68 ± 0.94
86.47 ± 1.90
72.00 ± 2.23
85.83 ± 1.83
Appendix
Table 7: Impact of differential privacy on federated concept-based models under different privacy budgets ε . We report task accuracy and concept accuracy (mean ± std.) for representative model instantiations.
Dataset
ϵs
C 2 BM
CGM
GD (%)
Asia
∞
80.4±2.3
77.0±3.9
0
8
80.4±2.3
77.0±3.9
0
4
80.4±2.3
77.0±3.9
0
2
80.2±1.9
77.9±1.7
3.6±3.6
1
79.8±2.4
77.8±1.5
21.4±11.3
Alarm
∞
72.6±0.9
71.1±0.6
0
Appendix
Table 8: Edge-level structural privacy. Task accuracy (%) for C 2 BM and CGM and graph disagreement (GD, %) with respect to the non-private aggregate under different structural privacy budgets ϵs .
# Clients
K=10
K=20
K=50
K=100
Models
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
OpaqNN
80.61 ± 1.98
91.63 ± 0.70
79.83 ± 0.99
90.80 ± 0.40
80.37 ± 1.06
90.03 ± 0.99
79.94 ± 1.15
90.30 ± 0.82
C 2 BM
80.14 ± 2.02
91.23 ± 0.81
78.16 ± 3.11
87.81 ± 2.32
80.35 ± 0.72
86.44 ± 2.17
75.38 ± 3.98
86.10 ± 2.42
CBM
79.91 ± 1.98
91.52 ± 0.83
79.25 ± 1.12
91.09 ± 0.61
79.54 ± 1.02
90.94 ± 0.50
78.32 ± 1.33
90.54 ± 0.55
CEM
80.58 ± 2.04
91.40 ± 0.72
80.58 ± 1.11
88.59 ± 1.63
79.92 ± 1.31
87.92 ± 2.01
78.80 ± 1.80
86.19 ± 1.91
CGM
79.33 ± 2.20
91.05 ± 0.89
79.30 ± 1.74
87.76 ± 1.83
78.96 ± 1.77
85.86 ± 2.10
78.84 ± 1.85
86.41 ± 2.14
Appendix
Table 9: Scaling with the number of clients. Task accuracy and concept accuracy (mean ± std.) as the total number of clients K increases, while keeping the overall amount of data fixed.
Model
Freezing
Asia
Sachs
Alarm
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
C 2 BM
Yes
80.35 ± 2.46
90.55 ± 1.08
76.20 ± 2.38
72.22 ± 1.91
73.14 ± 0.80
90.23 ± 0.31
No
80.52 ± 2.24
90.47 ± 1.29
76.55 ± 2.20
72.73 ± 1.59
72.68 ± 1.28
89.12 ± 0.84
CEM
Yes
80.50 ± 1.90
89.77 ± 2.33
75.97 ± 2.92
73.03 ± 1.55
73.42 ± 0.50
90.10 ± 0.63
No
80.27 ± 1.99
90.07 ± 1.92
76.34 ± 2.58
73.33 ± 1.38
72.79 ± 0.50
89.87 ± 0.71
Appendix
Table 10: Effect of selective parameter freezing ( Asia/Sachs/Alarm ). Task and concept accuracy (mean ± std.) across 5 seeds, comparing standard training (No) vs. selective freezing of unsupervised modules (Yes).
Model
Freezing
Insurance
Hailfinder
SIIM Pn.
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
Task Acc.
Concept Acc.
C 2 BM
Yes
74.27 ± 2.67
76.08 ± 1.37
72.83 ± 2.35
57.81 ± 1.61
68.80 ± 2.87
67.68 ± 1.51
No
74.47 ± 3.01
76.23 ± 1.44
71.70 ± 1.81
56.61 ± 1.45
65.90 ± 2.99
64.91 ± 3.07
CEM
Yes
70.97 ± 0.86
77.16 ± 0.80
70.15 ± 4.36
60.13 ± 1.43
66.33 ± 3.06
64.90 ± 1.80
No
72.62 ± 2.27
76.92 ± 1.39
70.30 ± 4.24
59.34 ± 1.51
66.91 ± 2.90
65.88 ± 3.24
Appendix
Table 11: Effect of selective parameter freezing ( Insurance/Hailfinder/SIIM ). Task and concept accuracy (mean ± std.) across 5 seeds, comparing standard training (No) vs. selective freezing of unsupervised modules (Yes).
Figure 15: Sensitivity analysis on the concept loss weight γ . Task accuracy (%, top row) and concept accuracy (%, bottom row) as a function of γ for the C 2 BM and CEM instantiations on Asia , Alarm , and SIIM-Pneumothorax , averaged over three runs. Performance is largely stable across values of γ . On SIIM-Pneumothorax , increasing γ yields a slight improvement in concept accuracy, accompanied by a small decrease in task accuracy, reflecting the expected trade-off between concept fidelity and downstream prediction.
Asia
Alarm
Model
10%
30%
60%
90%
No arr.
10%
30%
60%
90%
No arr.
CEM
79.3±1.3
79.9±1.2
79.8±0.9
77.8±2.5
67.7±12.9
73.6±0.2
72.6±0.7
70.4±1.4
70.3±1.5
55.1±17.3
C 2 BM
79.6±1.3
79.8±1.0
79.8±0.9
73.8±6.0
63.5±11.5
73.7±0.4
72.8±0.1
72.5±1.0
71.0±2.3
54.7±17.6
Appendix
Table 12: Sensitivity to client arrival timing. Task accuracy (mean ± std.) for F-CMs instantiated with CEM and C 2 BM when new clients arrive after different communication rounds. “No arr.” denotes the setting without client arrival. Results are reported over three random seeds.
Coverage range
Method
Asia
Alarm
S.C. Cov.
T.
S.C. Cov.
T.
<51%
S-F-CMs (CEM)
45.8±4.2
71.3±7.8
37.0±3.3
36.5±1.6
S-F-CMs (C 2 BM)
45.8±4.2
66.7±12.3
37.0±3.3
38.5±4.5
51 – 75%
F-CMs (CEM)
62.5
76.1±3.1
65.8±3.9
49.7±12.3
F-CMs (C 2 BM)
62.5
80.2±1.1
65.8±3.9
47.6±13.4
>75%
F-CMs (CEM)
100.0
80.3±1.1
100.0
73.7±0.7
Appendix
Table 13: Sensitivity to concept-space expansion. Structural concept coverage (S.C. Cov., %) and task accuracy (T., %) as additional concepts become available during federation growth. The coverage-range column denotes the target fraction of concepts from the ground-truth graph available to the federation. Results are reported over three random seeds.
Figure 16: Label accuracy (%) under inter-client semantic disagreement. Results for two representative F-CM instantiations, the bipartite F-CMs (CEM) and the graph-based F-CMs (C 2 BM) , under increasing fractions of within-client concept-label swaps. Each row corresponds to a disagreement regime: 30% of concepts and 30% of clients selected (top), or 60% of concepts and 60% of clients selected (bottom). For each selected concept, the corresponding fraction of clients is chosen, and within them a fraction of the concept labels (x-axis: 30%, 60%, 90%, 100%) is swapped, simulating conflicting concept semantics across clients. Each column corresponds to a different dataset.
Dataset
Model
Clean ref.
μ=0.3
μ=0.6
μ=0.9
Alarm
C 2 BM
71.8±0.9/89.9±0.5
72.4±0.7/89.7±0.2
72.6±0.3/89.1±0.1
70.1±1.1/85.1±0.9
CEM
69.9±2.8/90.2±0.3
62.3±9.6/90.0±0.4
58.1±9.5/89.4±0.2
56.8±9.2/86.8±0.3
Asia
C 2 BM
81.0±1.6/91.2±0.6
81.1±1.7/91.0±0.8
81.1±1.9/90.7±0.6
81.2±1.8/89.4±0.7
CEM
81.3±1.4/91.0±0.7
81.2±1.5/90.9±0.6
81.3±1.6/90.7±0.8
81.0±1.5/89.2±1.0
Appendix
Table 14: Robustness to missing concept annotations. Task/concept accuracy (%, mean ± std.) under increasing probabilities μ of masking available concept annotations. Results are averaged over three random seeds.
Dataset
Model
Clean ref.
η=0.3
η=0.6
η=0.9
Alarm
C 2 BM
71.4±0.9/89.9±0.5
71.7±0.5/87.9±0.4
72.1±0.9/46.7±2.4
51.7±9.6/14.9±2.2
CEM
69.9±2.8/90.2±0.3
73.1±0.6/89.1±0.6
71.6±1.3/47.1±2.7
50.5±12.2/7.2±0.7
Asia
C 2 BM
81.0±1.6/91.2±0.6
81.3±1.3/89.3±1.0
64.1±8.4/30.6±9.4
63.1±7.4/24.9±5.9
CEM
81.3±1.4/91.0±0.7
81.7±1.4/88.6±1.0
81.3±1.2/16.1±1.8
72.5±7.2/15.5±1.8
Appendix
Table 15: Robustness to noisy concept annotations. Task/concept accuracy (%, mean ± std.) under increasing probabilities η of corrupting available concept annotations. Results are averaged over three random seeds.
Figure 17: Robustness of DAG aggregation to client graph perturbations. DiffPairs (lower is better) between the aggregated DAG and the ground-truth DAG as a function of % client alteration and per-client graph alteration.
Dataset
Method
0%
30%
60%
90%
Asia
F-CMs (C 2 BM)
80.35 ± 2.16
80.41 ± 1.80
80.25 ± 1.98
80.39 ± 2.51
S-F-CMs (C 2 BM)
59.73 ± 20.19
59.99 ± 20.15
59.99 ± 20.15
60.83 ± 21.46
SIIM-Pneumothorax
F-CMs (C 2 BM)
68.80 ± 2.51
68.03 ± 3.34
68.62 ± 2.62
67.06 ± 2.90
S-F-CMs (C 2 BM)
62.68 ± 12.67
62.13 ± 1.01
62.14 ± 1.02
63.07 ± 2.58
Appendix
Table 16: Downstream robustness to graph misspecification. Task accuracy (%, mean ± std.) for C 2 BM under the same perturbation protocol as Figure 17 , with graph alteration fixed at 60% and varying fractions of corrupted clients. Although graph quality degrades at higher corruption levels, downstream task accuracy remains largely stable, especially for F-CMs.
Figure 18: Ablation of graph aggregation strategies under moderate perturbations. DiffPairs (lower is better) for different aggregation rules under localized, static federated, and dynamic federated visibility, with graph alteration p=0.3 and 30% corrupted clients.
Figure 19: Ablation of graph aggregation strategies under stronger perturbations. DiffPairs (lower is better) for different aggregation rules under localized, static federated, and dynamic federated visibility, with graph alteration p=0.6 and 60% corrupted clients.
Consensus-based aggregation
Data-size-based aggregation
Dataset
CA
GA =0.3
GA =0.6
GA =0.9
GA =0.3
GA =0.6
GA =0.9
Asia
0.3
0.0±0.0
0.0±0.0
0.0±0.0
0.0±0.0
0.0±0.0
0.0±0.0
0.6
0.0±0.0
0.0±0.0
0.3±0.6
0.0±0.0
0.0±0.0
1.0±1.7
0.9
0.0±0.0
0.3±0.6
14.3±1.2
0.0±0.0
1.0±1.0
14.3±1.2
Alarm
0.3
1.7±1.5
1.7±1.5
1.7±1.5
1.7±1.5
1.7±1.5
1.7±1.5
0.6
2.0±1.0
2.7±0.6
7.0±2.7
2.3±1.5
7.7±2.3
26.7±13.1
Appendix
Table 17: Consensus-based client weighting under graph perturbations. DiffPairs (lower is better) for consensus-based and data-size-based graph aggregation under varying client alteration (CA) and graph alteration (GA).
Figure 20: Multimodal F-CMs overview . At each round t , multimodal F-CMs follow the same steps as standard F-CMs: graph aggregation, dynamic architecture adaptation, and training. To handle heterogeneous input modalities, the shared CM includes, for each modality, a modality-specific encoder and a modality-alignment module, shown as the blue and yellow modules in the figure. The modality-specific encoder produces modality-specific embeddings, while the modality-alignment module maps them into a common latent space. Each client then processes its local inputs using only the encoder and alignment module associated with its modality, obtaining a latent representation that is then used to predict concepts and the task label.
Cent.
S-F-CMs
F-CMs
Task Acc.
Concept Cov.
Task Acc.
Concept Cov.
Task Acc.
Concept Cov.
OpaqNN
72.9 ± 0.5
100
58.4 ± 2.6
52.0 ± 10.2
70.3 ± 2.8
100
C 2 BM
73.3 ± 0.6
100
58.4 ± 2.5
60.0 ± 8.2
66.1 ± 2.7
100
CBM
71.3 ± 0.1
100
55.4 ± 3.0
52.0 ± 10.2
63.3 ± 1.4
100
CEM
73.2 ± 0.3
100
55.7 ± 3.1
52.0 ± 10.2
64.8 ± 2.0
100
CGM
73.3 ± 0.4
100
60.0 ± 2.9
60.0 ± 8.2
62.3 ± 1.9
100
Appendix
Table 18: Task accuracy (%) and concept coverage (%) with respect to the ground-truth concepts causally relevant for the task.
Dataset
Mean ζt2
Min. ζt2 (round)
Max. ζt2 (round)
Mean cosine sim.
Asia
0.0194
0.0186 (30)
0.0203 (1)
0.949±0.013
Alarm
0.0288
0.0287 (30)
0.0290 (1)
0.883±0.007
Appendix
Table 19: Empirical alignment mismatch. Empirical mismatch ζt2 between the module-wise aggregated direction and the gradient of the pooled objective over the first 30 communication rounds. The round attaining the minimum and maximum mismatch is reported in parentheses. Results are aggregated over three random seeds.
Cooperative inference across independently deployed machine learning models is increasingly desirable in distributed environments, as there is a growing need to leverage multiple models while keeping their data and model parameters private. However, existing cooperative frameworks typically rely on sharing input data, model parameters, or a common encoder, which limits their applicability in privacy-sensitive or cross-organizational settings. To address this challenge, we propose Consensus Embedding-based Federated Inference (CE-FI), a framework that enables pretrained models to cooperate at inference time without sharing model parameters or raw inputs and without assuming a common encoder. CE-FI introduces two components: a Consensus Embedding (CE) layer that maps heterogeneous intermediate representations into a common embedding space, and a Cooperative Output (CO) layer that produces predictions from these embeddings. Both layers are trained using shared unlabeled data only, so the cooperative stage does not require additional labeled data. Experiments on image classification benchmarks -- CIFAR-10 and CIFAR-100 -- under diverse non-IID conditions show that CE-FI consistently outperforms solo inference and performs comparably to conventional methods that require stronger sharing assumptions. Additional evaluations on text and time-series tasks indicate applicability beyond image classification, although performance depends on the ensemble strategy. Further analysis identifies representation alignment as the primary bottleneck.
Machine learning has become integral to medical research and is increasingly applied in clinical settings to support diagnosis and decision-making; however, its effectiveness depends on access to large, diverse datasets, which are limited within single institutions. Although integrating data across institutions can address this limitation, privacy regulations and data ownership constraints hinder these efforts. Federated learning enables collaborative model training without sharing raw data; however, most methods rely on complex architectures that lack interpretability, limiting clinical applicability. Therefore, we proposed a federated RuleFit framework to construct a unified and interpretable global model for distributed environments. It integrates three components: preprocessing based on differentially private histograms to estimate shared cutoff values, enabling consistent rule definitions and reducing heterogeneity across clients; local rule generation using gradient boosting decision trees with shared cutoffs; and coefficient estimation via ℓ1-regularized optimization using a Federated Dual Averaging algorithm for sparse and consistent variable selection. In simulation studies, the proposed method achieved a performance comparable to that of centralized RuleFit while outperforming existing federated approaches. Real-world analysis demonstrated its ability to provide interpretable insights with competitive predictive accuracy. Therefore, the proposed framework offers a practical and effective solution for interpretable and reliable modeling in federated learning environments.
Ke Wan, Kensuke Tanioka, Toshio Shimokawa
Department of Medical Data Science Wakayama Medical University · Department of Biomedical Sciences and Informatics Doshisha University · Department of Biostatistics Wakayama Medical University
Recent data protection laws have accelerated the adoption of Federated Learning (FL) for privacy-preserving decentralized training. Nevertheless, increasing model sizes impose substantial computational demands on client devices, limiting FL applicability in resource-constrained settings. We introduce a novel multi-domain federated learning framework in which lightweight client-side proxy models collaborate with a server-side Foundation Model (FM) to learn new concepts without sharing private data. Our approach, EFFEKT, enables efficient server-side training of domain-specific LoRA adapters while preserving feature-space alignment between the FM and proxy extractors via novel bi-directional cross-distillation strategies. Experiments on multiple real-world datasets and deployments on low-power edge devices demonstrate improvements over state-of-the-art baselines in most considered domains while maintaining lightweight computation at the client side.
Matteo Caligiuri, Francesco Barbato, Pietro Zanuttigh +1
Department of Electrical & Computer Engineering, Northeastern University, Boston (MA), United States · Department of Information Engineering, University of Padua, Padua, Italy