šŸ“¢Privacy isn’t a fancy feature, it’s a fundamental right

That’s why ARPA is pushing the boundaries of privacy-preserving computation šŸ”

Our upcoming ZKML research explores how we can keep your data private while still proving every blockchain & AI output šŸ”„

Stay tuned as we https://t.co/scHiNcNlFi https://twitter.com/arpaofficial/status/1977750926061678880
ARPA’s last few months were šŸ”„

From expanding AVS operators, launching new chain support, to pushing the boundaries of ZKML research and university programs

Catch all the highlights in our Fall 2025 Progress Report šŸ‘‰ https://t.co/d0PLt2v1hy https://t.co/AAmgoXDjk3 https://twitter.com/arpaofficial/status/1978612855764406443
6/ Paper #1
Demonstrating the Accuracy of an AI Model through Zero-Knowledge Proof is Even Faster Than Direct Computation.

Validation is not recomputation.
Freivalds-style checks and related math let you verify outputs cheaper than rerunning the model.
Includes a worked path for https://twitter.com/arpaofficial/status/1980333067895206190
7/ Paper #2
ZK-SNARK Verifiable Machine Learning.

What zk-snarks are, how they’re built, and how to program with them.
Arithmetization, proof systems, commitments.
Paths using Halo2, EZKL, Circom, circomlib-ml, keras2circom.
A blueprint for wrapping ML pipelines in proofs

Read https://twitter.com/arpaofficial/status/1980333069497512045
8/ Under the hood for builders:
Convolution, ReLU, and dense layers expressed as circuits via sumcheck and GKR.
Quantization from float to fixed point for ZK circuits with care for precision.
Off chain proving paired with on chain verification for open attestations. https://twitter.com/arpaofficial/status/1980333071091327241
2025/10/21 08:25:10
Back to Top
HTML Embed Code: