Polyhedra announced the release of zkPyTorch, a groundbreaking compiler that transforms PyTorch models/ONNX models into efficient, verifiable zero-knowledge circuits. AI models compiled with zkPyTorch can now generate cryptographic proofs that the model ran correctly, resulting in a major step forward for AI systems, enabling verifiability through zero-knowledge Machine Learning (zkML).
“zkPyTorch gives AI agents an identity,” said Tiancheng Xie, co-founder of Polyhedra Network. “It’s a trusted and scalable way to guarantee the integrity of an AI agent — without rewriting your AI stack,” Xie added.
Making Proofs Practical for Machine Learning
Until now, bringing zero-knowledge proofs (ZKPs) to deep learning required bespoke models and custom logic. zkPyTorch removes that barrier, integrating directly with standard PyTorch workflows and outputting circuits ready for ZKP engines like Expander, the world’s fastest prover, created by Polyhedra.
Through a novel compilation pipeline — spanning structured graph preprocessing, ZK-friendly quantization, and multi-level circuit optimization — zkPyTorch converts real-world models into field-efficient circuits that preserve performance and accuracy.
Key Benchmarks:
- VGG-16 (15M parameters): ~2.2 seconds per image proof
- Llama-3 (8B parameters): ~150 seconds per token
Performance measured on single-core CPU using Expander backend.
Also Read: Holdstation Unveils AI Agent Launchpad with $2M Grant
AI You Can Trust, Without Compromising Confidentiality
zkPyTorch is built to facilitate verifiability of both open-source and proprietary models. It ensures that inference correctness is cryptographically verifiable and enables the proof and result to be shared publicly.
This unlocks powerful use cases where AI is responsible for critical decisions, actions, and predictions, including:
- Trustworthy AI Agent Identity: Users securely assign an identity to an AI agent they trust, allowing it to safely perform critical tasks. Proofs verify that the results are genuine, come from the expected AI agent, and remain protected from tampering.
- Finance & Healthcare: Share AI decisions without exposing sensitive data.
- Compliance & Governance: Prove fairness or constraint adherence without leaking logic.
Developer-First by Design
zkPyTorch integrates seamlessly into developers’ existing PyTorch workflows by:
- Accepting standard trained models (via ONNX export)
- Applying quantization optimized for ZKP execution
- Outputting proof-compatible circuits for immediate use in Expander or compatible provers
SDKs are available in Python and Rust, with full docs and example integrations to get developers up and running quickly.
Source: PRNewswire