Mitigation of Sybil-based Poisoning Attacks in Permissionless Decentralized Learning
Résumé
Decentralized learning enables collaborative machine learning with enhanced privacy by allowing participants to train models locally and share updates for aggregation instead of sharing raw data. However, such systems are vulnerable to poisoning attacks that may compromise the learning process. This threat becomes even more severe when combined with sybil attacks, where adversaries contribute numerous malicious updates with minimal effort, amplifying their impact. To overcome these challenges, particularly in the permissionless setup, we propose SyDeLP, a blockchain-enabled protocol for decentralized learning. SyDeLP integrates byzantine tolerant aggregation for poisoning mitigation with a novel Verifiable Delay Puzzle to counter sybil attacks requiring Proofs of Work to participate. Honest behavior is incentivized by dynamically reducing puzzle difficulty, decreasing the computational burden for honest nodes over time. Empirical evaluations conducted on two benchmark datasets across four types of poisoning attack demonstrate that SyDeLP consistently outperforms existing solutions in terms of poisoning resilience.
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
licence |