ML Security PapersRobust and Efficient Collaborative Learning
Abdellah El Mrini , Sadegh Farhadkhan , Rachid Guerraoui
Published on arXiv
2510.08311
Data Poisoning Attack
OWASP ML Top 10 — ML02
Key Finding
RPEL maintains Byzantine robustness and near all-to-all accuracy while reducing communication complexity from O(n²) to O(n log n) in decentralized settings.
RPEL (Robust Pull-based Epidemic Learning)
Novel technique introduced
Collaborative machine learning is challenged by training-time adversarial behaviors. Existing approaches to tolerate such behaviors either rely on a central server or induce high communication costs. We propose Robust Pull-based Epidemic Learning (RPEL), a novel, scalable collaborative approach to ensure robust learning despite adversaries. RPEL does not rely on any central server and, unlike traditional methods, where communication costs grow in $\mathcal{O}(n^2)$ with the number of nodes $n$, RPEL employs a pull-based epidemic-based communication strategy that scales in $\mathcal{O}(n \log n)$. By pulling model parameters from small random subsets of nodes, RPEL significantly lowers the number of required messages without compromising convergence guarantees, which hold with high probability. Empirical results demonstrate that RPEL maintains robustness in adversarial settings, competes with all-to-all communication accuracy, and scales efficiently across large networks.
Key Contributions
- RPEL: a serverless, Byzantine-robust collaborative learning protocol with O(n log n) communication complexity via pull-based epidemic gossip
- Convergence guarantees under adversarial conditions that hold with high probability
- Empirical demonstration that RPEL matches all-to-all communication accuracy while scaling efficiently across large networks
🛡️ Threat Analysis
RPEL is a Byzantine-fault-tolerant protocol defending against malicious participants who send arbitrary model updates to degrade the global model during training — the canonical ML02 threat in federated/collaborative learning settings. The paper explicitly targets 'training-time adversarial behaviors' and evaluates robustness in adversarial settings.