dagfed-chain

Scalable multihierarchical sharded blockchain solution for federated learning

Blockchain platform consists of two layers:

• Mainchain layer - based on DAG (Tangle) consensus protocol, for maximum throughput link
• Subchain layer - blockchain layer consisted of multiple sharded hyperledger fabric networks. Each shard does federated learning task (FedAvg) independently of other shards and uploads its global model to the mainchain

Solution discussion

Systems heterogeneity

Be able to handle various client hardware, especially resource-constrained edge devices

• Find appropriate model architecture, fit for resource-constrained devices

  • MobileViT (https://arxiv.org/pdf/2110.02178.pdf)
  • MobileNet (https://arxiv.org/pdf/1704.04861.pdf)
  • DeIT (https://arxiv.org/pdf/2012.12877.pdf)

• Find alternate algorithm for federated learning, other than FedAvg, that takes into account the systems heterogeneity - MetaFed(https://arxiv.org/pdf/2206.08516v3.pdf)

Data heterogeneity

Be able to handle non-I.I.D. datasets, across all valid clients, in order not to lose generality

• Find appropriate model architecture that takes into account non-IID data - Transformer (https://arxiv.org/pdf/2106.06047.pdf)
• Find algorithm other than FedAvg, which takes into account unbalanced datasets - https://arxiv.org/pdf/2105.10056v2.pdf

Privacy preservation

TODO