- Polyhedra has launched zkPyTorch, allowing developers to create verifiable machine-learning models using standard PyTorch code.
- The tool converts PyTorch models into zero-knowledge proof circuits without requiring cryptographic expertise.
- zkPyTorch supports privacy, model integrity, and fast-proof generation for large-scale AI applications like Llama-3.
Polyhedra has launched zkPyTorch, a tool designed to make zero-knowledge machine learning (ZKML) accessible to developers using the PyTorch framework. The compiler translates standard PyTorch code into zero-knowledge proof (ZKP) circuits, enabling secure, verifiable AI inference without exposing sensitive model data or internal operations.
Introducing zkPyTorch: Making Zero-Knowledge ML Accessible! Polyhedra bridges PyTorch & ZK Proofs. Now AI devs can build verifiable, private ML models using standard PyTorch code – no crypto expertise needed!
✅ Prove model execution correctness
🛡️ Protect model IP & sensitive… pic.twitter.com/ZVD2MnrwGe— Polyhedra (@PolyhedraZK) June 5, 2025
zkPyTorch allows developers to preserve AI output integrity while keeping their intellectual property intact. The tool uses cryptographic techniques to show that a model is executed correctly without revealing model parameters or training data. Since the library requires no background in cryptography, the barrier to adoption to areas where sensitive information is handled, such as healthcare and finance, is lowered.
How zkPyTorch Optimizes AI Proof Generation
The framework incorporates three key modules to operate the complex machine-learning computations. Model preprocessing with ONNX format, to standardize ML graph representation is its beginning.
The second module teaches ZKfriendly quantization, replacing floating point operations with finite field arithmetic. The last module weaves circuit optimization wherein lookup tables can handle batch processing and nonlinear operations efficiently.
The model structure uses Directed Acyclic Graphs (DAGs) as its basis. Each operation, such as matrix multiplication or ReLU, is encoded as a node, resulting in elegant and efficient translation into ZKP circuits. This design works with even complex models like transformers and ResNets that are widely used in large language models (LLMs).
In addition, parallel circuit execution and FFT-based optimizations for convolutional layers accelerate proof generation. Multi-core hardware allows developers to increase throughput and decrease latency. Benchmarks show zkPyTorch processes Llama-3’s 8 billion parameters at a rate of roughly 150 seconds per token, retaining 99.32% cosine similarity with original outputs.
Real-World Use Cases and Future Directions
Polyhedra sees immediate applications in verifiable Machine Learning-as-a-Service (MLaaS), where cloud-based models can now provide cryptographic proof of correct inference. AI developers retain model confidentiality, while users gain assurance that outputs are valid. zkPyTorch also enables secure model valuation, giving stakeholders a trusted way to assess model performance without risking exposure to proprietary data.
The tool integrates with Polyhedra’s EXPchain, bringing verifiable ML to blockchain environments. This opens the door for AI-powered decentralized applications with on-chain validation.In a recent interview, Polyhedra’s founder noted, “We are targeting the multi-billion dollar market of zero-knowledge proofs. Our aim is to become the foundation layer of blockchain technology and expand the usage of zk proofs to sectors such as banking and other privacy-sensitive areas.”
