Zero-knowledge admission proposed for private swarm learning
Original title: Robust and Private Swarm Learning via Zero-Knowledge Validator Admission and Consensus Aggregation
Authors say the design holds 94% of clean CIFAR-100 accuracy under a gradient-scaling attack that drops FedAvg to 2.1%.
The paper proposes a swarm-learning framework that admits validators only after a zero-knowledge attestation that a reference model ran correctly on private data, without revealing raw samples. Attested outputs build a label-free class profile used to assemble a more balanced validator pool. Admitted validators score updates and use Tendermint-style Byzantine consensus to set a gradient mask and quality-based aggregation weight. A split-learning-style audit lets peers recheck masks. On four non-IID benchmarks it matches FedAvg on clean data and retains most accuracy under a gradient-scaling attack, the authors report.
Unconfirmed: Attack-resistance and pool-balance figures are author-reported and not independently confirmed here.
- Validator pool total-variation distance to uniform is reported as 0.025.
- Oracle and random selection are reported at 0.021 and 0.21.
- FashionMNIST retained accuracy is reported above 99%, versus 11% for FedAvg.
- Quality-derived weights replace dataset-size weighting.
- The audit avoids an end-to-end zero-knowledge proof.
From the original article
1 min read at arxiv.orgSwarm Learning improves resilience and removes the need for a single trusted coordinator by distributing model aggregation and validation across participating nodes.
By Marco Arazzi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.