Four years after introducing the concept of zero-knowledge machine learning, Berkeley RDI and Polyhedra have unveiled a production-ready system that could redefine how trust and transparency are built into artificial intelligence.
The new zkML technology, announced today and shared with crypto.news, enables developers to prove the correctness of AI outputs without exposing sensitive underlying data or models, according to a company press release.
zkML
At its core, zkML applies zero-knowledge proofs to machine learning. ZKPs are a cryptographic technique that allows one party to prove a statement is true without revealing the data behind it.
This approach addresses trust concerns in AI, which often involve “black box” systems that lack transparency. With zkML, users can confirm that AI systems work as intended while maintaining privacy and compliance.
From Research to Reality
The concept of zkML was first introduced in 2020 by Jiaheng Zhang, Chief Scientist at Polyhedra, along with Berkeley researchers Yupeng Zhang and Dawn Song. At the time, zkML was purely theoretical due to the high computational demands of ZKP systems, according to the release.
Today, advances in zero-knowledge technology, such as Polyhedra’s Expander proof system, have made it practical to deploy zkML in real-world scenarios.
How zkML will be used
Beyond verifying AI outputs, zkML has the potential to transform how AI systems manage privacy and accountability. It facilitates data origin verification, ensuring the authenticity and traceability of AI training data, while enabling authenticated data labeling to verify that labeled data remains accurate and unaltered. Additionally, zkML allows for training process validation, proving that AI models were trained according to strict protocols.
Polyhedra envisions zkML playing a significant role in combining AI with blockchain technology. This could support decentralized AI ecosystems, secure model deployment, and privacy-focused applications.
As zkML evolves, its backers see it as a tool to build trust in AI applications without compromising on privacy or security.
According to the release, Polyhedra and Berkeley RDI plan to expand zkML’s capabilities further, making the technology accessible to developers with minimal expertise in cryptography.
Read the full article here