Swarm Learning improves resilience and removes the need for a single trusted coordinator by distributing model aggregation and validation across participating nodes. However, it also delegates the validation of model updates to peers whose private data may be biased, unrepresentative, or adversarial, and existing Swarm Learning protocols generally do not attest that data, leaving decentralized validation as an attack surface. We present a robust, privacy-preserving Swarm Learning framework that makes verifiable data quality the condition for admission to the learning task. A validator joins only after a zero-knowledge attestation proves, without revealing any raw sample, that a reference model ran correctly on its private data. The attested outputs also yield a label-free class profile, so the network assembles a collectively IID validator pool from individually skewed validators as accurately as an oracle with the true labels (total-variation distance to uniform 0.025, against 0.021 for the oracle and 0.21 for random selection). Admitted validators score client updates and, through a Tendermint-style Byzantine Fault Tolerant consensus, agree on a gradient mask and a quality-derived aggregation weight for each submission, replacing dataset-size weighting with a quality-aware signal. Adapting the cut-layer idea of Split Learning, a challenge-based audit lets peers re-verify that a mask was computed on the attested data without an expensive end-to-end zero-knowledge proof. Across four benchmarks under non-IID partitions the framework matches FedAvg on clean data; under a gradient-scaling attack that collapses FedAvg to random chance, it retains 94% of clean accuracy on CIFAR-100 and over 99% on FashionMNIST, where FedAvg achieves only 2.1% and 11%. The result is a private, robust, and auditable foundation for decentralized learning.
Figures & tables
Acronym
Full Term
FL
Federated Learning [ 26 ]
SL
Swarm Learning [ 33 ]
ZKP
Zero-Knowledge Proof [ 11 ]
BFT
Byzantine Fault Tolerant [ 8 , 7 ]
MPC
Multi-Party Computation [ 38 ]
R1CS
Rank-1 Constraint System [ 13 ]
Table 1: Key Acronyms and Their Full Terms
Figure 1: End-to-end architecture of the robust and private Swarm Learning framework, shown across the four numbered stages of one round. (1) Clients train local models and send them to the Model Pool. Each model is also uploaded to IPFS, and its hash is registered on-chain with the Coordinator SC, so every submission is authenticated and cannot be altered. (2) A client seeking validator status runs the reference model on its private data and sends the on-chain Verifier SC a zero-knowledge proof that it ran the model correctly. The Verifier SC verifies the proof, and the client is admitted only if its outputs also pass a quality gate that the network applies off-chain, after which the mix of classes the client’s data contains (its class profile) is recorded. The admitted validators are picked so that, together, their data is balanced (IID) across all classes. (3) Each validator scores the pooled client models and builds a gradient mask for every one. A consistency check then removes any validator whose mask its peers cannot reproduce. (4) The validators run a BFT consensus to agree on the masks and combine the client updates using weights set by that consensus. The result is the next global model, which is committed on-chain and stored on IPFS.
Figure 2: Challenge-based mask verification as an interaction between the challenged validator and the peer auditors, split by the cut layer ℓ . The validator re-runs the model head f≤ℓ on its private admission batch and exposes only the smashed activations aℓ , the single quantity that crosses the cut layer. Peers complete the forward pass and remove the validator if the recomputed logits disagree with its committed logits; otherwise they back-propagate to layer ℓ , reconstruct the gradient G^ , and retain the validator only when at least 95% of its kept mask coordinates stay above the published threshold τ with none scaled down.
Phase
Time
Space
Communication
Admission
O~(bF)
O(bF)
O(1)
Eval + masking
O(M(nF+d))
O(d)
O(Md)
Challenge audit
O(P(bF+d))
O(d)
O(b∣Aℓ∣)
Consensus
O(Vd)
O(Vd)
O(V2)
Table 2: Asymptotic cost of the validator protocol. Admission is paid once per validator; evaluation, masking, and consensus are per round, and the audit is per challenge. Here V=3f+1 validators, M pooled models, P peers, d parameters, n test samples, b batch, F forward-pass cost, S=Θ(bF) circuit size, and ∣Aℓ∣ the cut-layer activation size.
Scale factor
2 ×
5 ×
10 ×
50 ×
Accuracy retained
99.5%
99.4%
11%
21%
Table 3: Scale-factor sweep for the gradient-scaling attack on FashionMNIST (20% attackers, α=0.3 ). Accuracy retained is the best accuracy under attack relative to the no-attack FedAvg baseline; higher means the attack did less damage. The undefended baseline tolerates small factors and collapses at 10 × , the value we adopt as the default.
N=10,V=3
N=20,V=5
N=30,V=6
Dataset
α
FedAvg
p-iid l-niid
p-iid l-iid
FedAvg
p-iid l-niid
p-iid l-iid
FedAvg
p-iid l-niid
p-iid l-iid
F-MNIST
0.1
86.5
85.3
87.0
88.6
89.0
88.5
87.6
84.9
85.4
0.3
89.7
89.8
89.4
89.5
90.2
90.9
89.6
89.7
90.8
1.0
91.4
92.0
91.8
90.9
91.3
91.0
90.7
91.1
91.2
IID
91.8
NA
92.0
91.6
NA
92.0
91.5
NA
91.5
CIFAR-10
0.1
63.7
61.9
64.0
63.7
68.7
76.1
41.4
52.9
51.4
Table 4: FedAvg baseline vs. Mask Consensus (Ours). Best accuracy (%). Ours variants: pool-iid/local-non_iid, pool-iid/local-iid. Bold = winner per row-topology.
Figure 3: Per-client aggregation weight over 61 rounds (CIFAR-10, α=0.1 , N=10 , V=3 ). Solid blue: consensus-derived weight; dashed grey: static FedAvg weight proportional to local dataset size. Blue shading marks rounds where a client receives more influence than its data size would warrant; red shading marks the opposite.
Figure 4: Effect of Dirichlet α on our framework (CIFAR-10, N=20 , V=5 ). Left: best test accuracy across 250 rounds. Right: communication rounds needed to first reach 70% accuracy; α=0.05 never reaches this threshold and is omitted.
Figure 5: Effect of validator count V on our framework (CIFAR-10, N=10 , α=0.3 ). The total training data (40 000 samples) and validator pool (10 000 samples) are held constant; increasing V reduces the number of training clients. Left: best test accuracy. Right: communication rounds to first reach 70% accuracy.
Figure 6: Accuracy retained under gradient scaling attack (scale factor 10 × , α=0.3 ). Each subplot shows accuracy retained relative to the no-attack FedAvg baseline (dashed grey line at 100%) as a function of attacker fraction. Rows: datasets; columns: network topologies. Shaded regions denote ± 1 standard deviation over three seeds. Our method (solid blue) maintains near-baseline accuracy across most configurations where FedAvg (dashed red) collapses to random chance.
Selection strategy
Test-set TV to uniform
Imbalanced population (all 60)
0.160
Random (20 validators)
0.207
Oracle, true labels (20)
0.021
Score, label-free (20)
0.025
Table 5: Composing an IID validator test set on FashionMNIST (population of 60 individually-skewed validators, 20 selected). TV is the total-variation distance of the selected test set’s class distribution from uniform; lower is more IID. The label-free score matches the label-based oracle and far outperforms random selection.
Figure 7: Label-free composition of an IID validator pool (FashionMNIST, population of 60 skewed validators). Left: as validators are selected, the combined test set’s distance from a uniform (IID) class distribution; the label-free score (blue) tracks the label-based oracle (dashed) almost exactly and far outperforms random selection (red). Right: the resulting test-set class distribution at twenty validators is flat under the score but lopsided under random, which over-represents common classes and starves rare ones.
Phase (cadence)
Time
Measured
Footprint
Admission (once / validator)
O~(bF)
84 s prove, 3.0 s verify
60.5 GB host, 879 KB proof
Eval + masking (per round)
O(M(nF+d))
none
O(d) state, O(Md) comm.
Challenge audit (per challenge)
O(P(bF+d))
69 ms
12.5 KB exposed
Consensus (per round)
O(Vd)
none
O(V2) messages
Table 6: Per-phase cost of the framework, asymptotic and measured in one place. The time column is recalled from Table 2 , where the full space and communication bounds are derived, and is paired here with the measured cost on the FashionMNIST reference model. Admission is paid once per validator and amortized over all later rounds; evaluation, masking, and consensus recur each round; the challenge runs per audit. An entry of none marks a phase with no cryptographic overhead, the standard forward and backward passes or Byzantine Fault Tolerant messaging, which we do not benchmark separately. One-time admission artifacts are 97 s setup, an 18.4 GB proving key released after admission, and an 8.6 MB verifying key. Symbols follow Table 2 : b batch size, F a reference-model forward pass, M pool size, n test-set size, d parameters, V validators, P challenge peers.
Dataset
Cut layer
SSIM
Honest
Adv.
Audit
FashionMNIST
features
0.876
1.000
0.080
ok
CIFAR-10
layer2
0.876
1.000
0.076
ok
layer3_pool
0.494
1.000
0.080
ok
layer3_res
0.288
0.762
0.324
weak
GTSRB
layer2
0.914
1.000
0.065
ok
layer3_pool
0.700
1.000
0.082
ok
Table 7: Cut-layer selection trades validator data privacy against audit strength. SSIM scores how well an inverter reconstructs the input from the exposed activations (lower is more private). Honest and Adv. are the mask-agreement rates of an honest and a fabricated validator; the audit is sound when honest agreement stays above the 0.95 acceptance threshold and adversarial agreement stays low. weak marks a cut too deep to audit reliably. Bold marks the cut used in our framework.
Department of AIML Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology Chennai, India · Assistant Professor, Department of AIML Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology Chennai, India · Professor, Department of AIML Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology Chennai, India