Home · Features
dfpn is the coordination layer under whichever detectors operators choose to run — verifiable AI built on a DePIN network of operator GPUs. Its features are the properties that make a decentralized verdict trustworthy: no single party can copy, buy, or quietly rewrite it. Everything below traces to the protocol design in the repo — no benchmark theatre.
Verifiable AI
Each assigned worker posts a SHA-256 commitment of its result plus a salt, then reveals after the commit window closes. The chain checks every reveal against its commitment, so a worker cannot copy another’s answer or wait to see the crowd before deciding.
A request is served by several independent workers running their own models on their own GPUs. A single wrong, offline, or malicious worker cannot move the verdict on its own.
The final verdict and confidence score are a reputation-weighted aggregation of revealed results — not a naive majority. Consistently accurate operators carry more weight; new or unreliable ones carry less.
Economic security
Workers post at least 5,000 DFPN; model developers post at least 20,000 DFPN per model version; challengers escrow 5% of the disputed reward. Floors scale with request fees so participation stays costly to fake.
Invalid results are slashed 10% of stake. Fraud or collusion is slashed 25–50% plus a temporary ban. Missed deadlines cost 1–3% plus reputation decay. The penalty fits the offense.
Rewards weight accuracy 50%, availability 25%, latency 15%, and consistency 10%. Stake weight enters the calculation but is capped, so capital alone cannot buy the verdict.
Open & composable (DePIN)
Any model developer can register a detection model on-chain with metadata, versioning, and a stake. The reference worker ships four models — face manipulation, AI-generated images, video authenticity, and voice cloning — but operators are never limited to them.
dfpn never sees inference. Operators contribute their own GPUs and choose their own algorithms; if one drops offline, the rest of the pool keeps serving requests. Detection capacity is decentralized physical infrastructure — it scales with whoever plugs in a GPU, with no central data center to capture or take down.
The reference client covers image, video, and audio — including face manipulation and voice cloning. Model output is standardized JSON with a verdict, confidence, and per-detection regions.
Audit & guardrails
Request, assigned workers, commitments, reveals, and the aggregated verdict are all anchored on Solana. Anyone can reconstruct how a piece of media was judged and which models contributed — there is no admin endpoint that silently rewrites a result.
Random assignment, diversity constraints on models and operators, and a challenge window before slashing make cartels expensive and slow. Diversity is treated as a security primitive, not a nice-to-have.
Dynamic fees, rate limits, priority fees, and per-epoch reward caps make it costly to flood the network with low-fee requests or to swarm it with cheap sybil identities.
Features are only as good as the architecture underneath them. Walk the on-chain programs, off-chain services, and the full lifecycle of a verdict.