Computation on data you cannot see. Verification of work you cannot read.
Dark Subnet is a pioneering Bittensor implementation that unlocks sensitive AI use cases (Healthcare, Finance, Privacy-Preserving GovTech) by combining Fully Homomorphic Encryption (FHE)/Zero Knowledge Proofs (ZKP)/Multi Party Computation (MPC) with a novel Honey Pot Verification mechanism.
Today, decentralized AI requires miners to see raw inputs. This prevents deployment in healthcare, finance, government, and enterprise environments where data exposure is unacceptable.
Miners compute directly on encrypted data. They never see inputs or outputs.
Using Pedersen commitments and Fiat-Shamir transforms, miners prove correct computation without revealing data.
Validator keys are distributed using Shamir Secret Sharing with threshold decryption. No single validator can decrypt sensitive results.
Verifiable AI work that remains completely private.
Validators inject encrypted Honey Pot Trap tasks to detect dishonest miners. Since miners cannot distinguish traps from real requests, correctness proves genuine encrypted computation.
The computational overhead of FHE becomes an inherent proof-of-work — intelligence cannot be faked cheaply.
---
Dark Subnet transforms Bittensor from a marketplace of intelligence into a marketplace of trusted intelligence.
It is the cryptographic trust layer for decentralized AI.
Testnet Guide: https://github.com/GodOfAgents/Dark-Subnet/blob/main/TESTNET_GUIDE.md
(Ran all tests in Docker)
docker run dark-subnet python -m pytest tests/ -v
Results: 36 passed
Crypto (ZKP/MPC): 26 tests
FHE Models: 6 tests
Synapse Protocol: 4 tests
$0/- no fundraising fully self funded