A secure model-lifecycle layer for federated and gossip learning in smart grid attack detection

Smart grids are large-scale cyber–physical systems in which distributed sensing, control, and monitoring generate sensitive operational data across multiple administrative domains. In this setting, centralized machine learning raises well-known concerns related to scalability, trust concentration, and single points of failure, while also making model management dependent on a single coordination point. Distributed learning paradigms such as Federated Learning (FL) and Gossip Learning (GL) offer a more suitable direction, since they preserve data locality and reduce the need for centralized data collection. However, they also introduce new problems related to artifact integrity, controlled access, provenance, revocation, and recovery in partially trusted environments. This paper introduces a secure lifecycle layer for model artifacts in smart-grid attack detection. Local updates, aggregated models, peer-exchanged models, and checkpoints are governed independently of the learning protocol. The layer combines authenticated encryption before off-chain IPFS storage, BBS + -signed manifests, HPKE key release through a Key Release Service (KRS), prospective revocation, ledger-backed audit, and checkpoint recovery. The same lifecycle semantics are instantiated in Blockchain-enabled Federated Learning (BFL), with auction-based coordination, and Blockchain-enabled Gossip Learning (BGL), which preserves direct peer exchange without a privileged aggregator. The evaluation covers prediction, robustness, lifecycle security, and cost. Federated configurations converge faster; gossip configurations remain operational under message loss. At N = 100 and R = 100 , collect-offers accounts for 81.9% of the measured BFL component sum, while median KRS latency rises from 11.5 to 551.6 ms as the authorized group grows. Single-run functional validations in BFL and BGL demonstrate externally triggered revocation, prospective key-release denial, authenticated checkpoint restoration, and resumed learning after training-time label flipping; inference remains local.