zkDaily - 2026-08-20

零知识证明zkDaily

ZKP Frontier Tracker 🎯

Thu

08.20

2026

Certified but Private: Scalable Zero-Knowledge Proofs for Neural Network Guarantees

Paper

https://arxiv.org/abs/2608.17070

zkMLCROWN

Zhong et al. proposed PANDA, a scalable system using ZKPs to prove robustness and fairness of neural networks without revealing private parameters. Its core contribution is a novel algorithm for proving linear relaxation bounds for non-linear activations, enabling lightweight proofs and polynomial scaling.

NotesNotes要点

PANDA verifies model robustness and fairness with hidden parameters via ZKPs.
Built on CROWN, a robustness certification framework.
A new algorithm certifies linear relaxation bounds for non-linear layers, yielding lightweight proofs.
Handles networks with 2.9M+ parameters: ~5 min to prove, ~10 s to verify.
Polynomially scalable, supporting networks 4 orders of magnitude larger than prior ZKP systems.
Offers a new path for privacy-preserving model audits.

Collected by @icerdesign

零知识证明zkDaily

Q&A Deep Dive 💬

Thu

08.20

2026

Which model properties does PANDA verify?

It verifies robustness and fairness of neural networks without revealing model parameters.

Which existing verification framework does PANDA build on?

It builds on CROWN, a robustness certification framework widely used for formal verification of neural networks.

What improvements does PANDA offer over prior zk robustness systems?

It supports networks with 2.9M parameters, proving in 5 minutes and verifying in 10 seconds, scaling to networks 4 orders of magnitude larger.

Prompted by @icerdesign