I am working on a Decentralized AI (DeAI) project where we want to train models on private user data without the data ever leaving the user's device. How are people using Zero-Knowledge Proofs (ZKPs) or Fully Homomorphic Encryption (FHE) on-chain to verify that a model was trained correctly without exposing the raw training dataset?
3 answers
This is the "Holy Grail" of 2024-2025 tech convergence. The most common approach right now is using zk-SNARKs to create a proof of computation. Essentially, the user’s device trains the model locally and then submits a cryptographic proof to the blockchain. The smart contract verifies this proof to confirm the training happened according to the rules, but it never sees the underlying data. FHE is still a bit too computationally expensive for most mainnets, but we are seeing specialized "Layer 2s for AI" that are optimized specifically for these heavy mathematical operations. It’s a massive step forward for medical and financial AI applications.
Do you think the current gas costs for verifying ZK-Proofs on Ethereum are too high for frequent model updates, or are you looking at cheaper Alt-L1s for this?
Check out "Federated Learning" protocols. They complement blockchain perfectly by keeping data local while the chain handles the global model aggregation and rewards.
Federated learning is definitely the way to go. Combining it with blockchain-based incentive layers ensures that users are actually paid for the data they contribute.
We are definitely looking at Layer 2 solutions or even dedicated app-chains. Verifying a complex ZKP on Ethereum Mainnet can cost upwards of $50, which is fine for a final model release but impossible for daily training cycles. We need a sub-cent verification cost to make decentralized learning commercially viable for small-scale datasets or consumer apps.