Berkeley RDI and Polyhedra have entered into a strategic partnership to unveil a groundbreaking Production-Ready zkML. This production aims to make a transformative leap in the convergence of AI and cryptographic verification. The organizations are now delivering a real-world application, four years after they pioneered the concept. The application aims to enable AI developers to use zkML without the need for specific expertise in zero-knowledge proofs.
The journey began in 2020 when Polyhedra’s research team, including Jiaheng Zhang, Dawn Song and Yupeng Zhang, published a paper in collaboration with Berkeley RDI. The article was titled “Zero Knowledge Proofs for Decision Tree Predictions and Accuracy,” which introduced a zkML, zero-knowledge machine learning. The concept of the article focused on building trust in AI by ensuring verifiable results while maintaining the privacy of underlying data and models.
Polyhedra prioritizes reliable systems to minimize human error in technological incidents. At its core, the zkML model allows developers to prove the accuracy of AI model predictions on a specific data sample without exposing sensitive information.
zkML Polyhedra: Promoting trust and transparency in AI
Using zero-knowledge proofs, a service provider can prove that a specific output was actually produced by running a certain model on an input. zkML technology provides an effective solution to one of AI’s most pressing challenges, namely trust and transparency. By enabling verifiable calculations, the zkML model ensures that the results generated by AI models are accurate and reliable. This approach addresses concerns about the opaque nature of AI, where unverified results can lead to incorrect decisions.
The organization’s CTO, Tiancheng Xie, shared, “We have spent the entire life of the company building systems that can operate without human intervention, are verified by mathematics, and are cryptographically secure.”
Applications of zkML extend beyond inference verification and include areas such as authentication of data provenance, accurate data labeling, and verification of AI training processes. This ensures that every step of the AI lifecycle meets strict integrity standards.
The zkML model eliminates high computational barriers, built on the innovative Expander-proof system, making production-ready solutions a reality. These advances not only strengthen trust in AI, but also ensure compliance with privacy regulations.
The future of verifiable AI
Dawn Song, president of Berkeley RDI, emphasizes: “Berkeley RDI and Polyhedra are setting a new standard for trust and transparency in artificial intelligence with innovative zero-knowledge machine learning (zkML) technology, a breakthrough approach that combines machine learning with cryptographic verification.”
This collaboration sets a new benchmark for transparency and accountability in AI. Song emphasized that the goal is to ensure that AI’s efficiency does not compromise trust and security.
Looking ahead, zkML is poised to transform the AI landscape by enabling secure deployment, decentralized ecosystems, and innovative applications. With zkML, Berkeley RDI and Polyhedra want to build a future where trust is at the core of AI innovation.