How Zero-Knowledge Proofs Are Making AI Agents Smarter and More Efficient: The Future of Collaborative Computation
Artificial Intelligence (AI) technology is increasingly everywhere. It features at the top of search engine results, is being used to write minor lines of code, can conduct research and cite sources, and some people are even trying to figure out how to put it in your dishwasher. While the need for having AI in your dishwasher might be debated, one thing that cannot be debated is that using this much computing power has definite costs.
The energy needed to train and operate AI systems is immense. Massive data centers must be built to process the information required to make an AI understand your request and give a coherent answer. By next year, these data centers are expected to collectively use as much energy as the entire country of Japan. While not every aspect of these centers are dedicated to AI, the problem of their energy consumption must be addressed all the same.
Complicating matters is the increasing standards of people using AI. The tendencies of AI to hallucinate, misstate information, or outright make things up is increasingly well known. Users demand proof that the information provided by AI models is accurate, based on trustworthy data, and correctly processed. Given the importance of certain tasks AI is now being trusted with, such as image recognition, guiding robots, and deciding how to drive cars, it is understandable that people might want proofs that the AI model is using sound data to operate. Naturally, this calls for more processing, more data, and more energy use.
However, there is a way to address both of these problems at the same time. Zero Knowledge Proofs (ZK proofs) are a powerful tool that allows users to confirm the accuracy of information while protecting privacy. Used properly, it can do so with remarkable efficiency.
ZK Proofs Explained
For those who need a refresher, ZK proofs are methods for proving to somebody that one party has particular information without just showing them that information.
A common illustrative example is that of “Ali Baba’s Cave.” Imagine that you are in a magic cave shaped like a ring with a tunnel leading into it. On the opposite side of the ring from the tunnel, and hidden by the cave wall, is a magic door that only opens with a magic word. Suppose you have a friend who wants proof that you know the magic word, but you don’t want them to actually hear what it is. How would you do it?
An answer is to prove that the only way to go around the cave without passing the entrance tunnel is by the door and the only way through the door is to know the password. Then, by doing just that, you prove that you have the password — even though your friend still doesn’t know what it is.
It might be further clarified with another story. Imagine that your friend is colorblind. He doesn’t entirely believe that two balls you have, a red one and a green one that are otherwise identification, are different. To prove that they can be distinguished, you have him put them behind his back and then show you one. If he switches them behind his back, you would know. After a few rounds of correcting, telling him if he switched the balls, he believes that you can tell them apart even though you never told him which one is green and which one is red.
These proofs have a number of applications, particularly in authentication, privacy preservation, and related areas. They may well be of use in addressing the problems of AI.
ZK meets AI
It may well be possible to apply ZK proofs to AI in a way that confirms the accuracy of the information provided efficiently, effectively, and in a way that protects the data used to train the AI model. By using these proofs, a user could quickly confirm that the data cited by the AI model really exists, without necessarily having direct access to that data.
Additionally, this could be done fairly efficiently, addressing issues of energy consumption. The ZK-SNARK proof, which stands for “succinct non-interactive argument of knowledge”, is particularly small as far as proofs like this go, and could be used to prove the accuracy of many kinds of information at a low computational cost. Where speed is more important than the economy of data use, ZK-STARKs, “scalable transparent argument of knowledge”, could be of great use.
By applying ZK proofs to the outputs of AI models, it may be possible to add a level of certainty to AI outputs without adding a need for even more energy intensive computation. Given the potentialities of AI and the problems currently facing it when it hallucinates, addressing this issue should be a primary concern for many working in the field. The possible solutions promised by ZK proofs should be taken seriously by all involved.
About ARPA
ARPA Network (ARPA) is a decentralized, secure computation network built to improve the fairness, security, and privacy of blockchains. The ARPA threshold BLS signature network serves as the infrastructure for a verifiable Random Number Generator (RNG), secure wallet, cross-chain bridge, and decentralized custody across multiple blockchains.
ARPA was previously known as ARPA Chain, a privacy-preserving Multi-party Computation (MPC) network founded in 2018. ARPA Mainnet has completed over 224,000 computation tasks in the past years. Our experience in MPC and other cryptography laid the foundation for our innovative threshold BLS signature schemes (TSS-BLS) system design and led us to today’s ARPA Network.
Randcast, a verifiable Random Number Generator (RNG), is the first application that leverages ARPA as infrastructure. Randcast offers a cryptographically generated random source with superior security and low cost compared to other solutions. Metaverse, game, lottery, NFT minting and whitelisting, key generation, and blockchain validator task distribution can benefit from Randcast’s tamper-proof randomness.
For more information about ARPA, please contact us at contact@arpanetwork.io.
Learn about ARPA’s recent official news:
Twitter: @arpaofficial
Medium: https://medium.com/@arpa
Discord: https://dsc.gg/arpa-network
Telegram (English): https://t.me/arpa_community
Telegram (Turkish): https://t.me/Arpa_Turkey
Telegram (Korean): https://t.me/ARPA_Korea
Reddit: https://www.reddit.com/r/arpachain/