ZK, AI, ZKML: Disruptive Technologies & Real-Life Use-Cases

Woman in a rice field next to the ocean teaching about zero-knowledge technology

In this article, we’ll be talking about two disruptive technologies that gained a lot of traction in the past year: Zero-Knowledge Proofs (ZKP) and Artificial Intelligence (AI). We’ll also look into the exciting possibilities that emerge from their combination, known as Zero-Knowledge Machine Learning (ZKML). After a short but hopefully crystal clear explanation of how they work, we’ll dive into some specific use cases where they could make a tremendous impact on real life applications.

Artificial Intelligence

OpenAI’s ChatGPT has been taking the world by storm 6 months ago by showcasing impressive results with natural language processing (NLP), and it is a great illustration of how a technology can reach massive adoption extremely quickly despite being part of the scientific landscape for so long. Although the machine learning algorithms have gotten better in the past couple years indeed, it’s mostly because storage, bandwidth and processing power have followed Moore’s law that it is now possible to produce something that seems human — and trigger the “wow that’s magic” reaction.

But whichever the reason for this “iPhone moment” (with Artificial Intelligence in lieu of Smartphones), the disruptive potential of deep learning, powered by multi-layered neural networks, is undeniable and can potentially impact several industries. Before we dive into exploring some of its potential real-life applications, we’ll just briefly clarify these terms which are often used interchangeably: Artificial Intelligence, Machine Learning, Neural Networks and Deep Learning, which sort of encompass one another. Deep Learning (DL) is actually a subfield of Machine Learning (ML), ML is a subfield of Artificial Intelligence (AI), and DL is based on Deep Neural Networks (DNN), which themselves are a subtype of Artificial Neural Network (ANN), an algorithmic structure mimicking the behavior of the human brain. So a nice way to visualize it is, AI is the external layer of an onion, and then, layer by layer, progressively come ML, ANN, DNN and DL.

Artificial Intelligence as an onion: ML > ANN > DNN > DL

OK, now that the theory is out of the way, let’s try and list some potential use cases that could benefit from AI assistance, focusing on the respective specificities of some of the biggest economic sectors: Health & Pharmaceuticals, Banking & Insurance, and Education & Professional Training.

Health & Pharmaceuticals

• Drug discovery: AI can improve researchers’ productivity to identify new drug candidates, optimize drug design, and predict drug efficacy and toxicity by using large datasets, advanced analytics, and DL.
• Clinical trials: it can also help design better clinical trials, recruit and monitor patients, analyze trial data, and report trial outcomes by using NLP, computer vision, and ML.
• Diagnosis and treatment: clinicians could use AI assistance to diagnose diseases, recommend treatments, and monitor patient outcomes by using image recognition, voice recognition, and recommendation systems.
• Digital therapeutics: AI could be trained to deliver personalized and evidence-based interventions to patients through digital platforms, such as apps, devices, or software. These interventions can help prevent, manage, or treat various conditions, such as diabetes, mental health, or chronic pain.
• Product intelligence: monitoring and analyzing the performance, safety, and quality of health care products and services could be enhanced by using sensors, NLP and ML. This can help generate insights and feedback for improvement, as well as detect and report adverse events or defects.
• Manufacturing and supply chain: This aspect is actually valid for most non-service businesses, not only pharma. AI can help manufacturers optimize production processes, reduce costs, improve quality, and ensure compliance by using sensors, robotics, and ML.

Banking & Insurance

• Customer service: banks and insurance companies can provide personalized and seamless services to customers through AI-based voice assistants, NLP-able chatbots, biometrics, and recommendation systems.
• Fraud detection: monitoring and prevention of fraudulent activities would benefit AI capacity to analyze large amounts of data and identify patterns and anomalies.
• Data analytics: more generally, any kind of processing of vast troves of data can benefit from the use of NLP, computer vision, and DL, in order to generate insights and value.
• Credit scoring: AI can help banks assess the creditworthiness of borrowers by using alternative data sources (e.g. public info gathered from social networks) and smart contracts.
• Risk management: some banks are already working on ways to optimize their risk models and portfolios by using advanced analytics and ML.
• Claims management: AI can help insurers automate and streamline the claims process by using NLP, computer vision, and DL to assess damage, verify coverage, and estimate repair costs.
• Underwriting and pricing: insurance companies can leverage AI-based tools to personalize and optimize their products and premiums by using alternative data sources, advanced analytics, and ML to segment customers, predict risk, and tailor offers.

Education & Professional Training

• Personalized learning: with its potential to adapt flexibly to different learners behaviors, AI can tailor the learning content, pace, and style to the individual needs, preferences, and goals by using adaptive learning systems, recommender systems, and chatbots.
• Assessment and feedback: AI can provide timely and constructive feedback to learners and educators, but also help evaluate the learning outcomes, progress, and gaps of each learner by using NLP, computer vision, and ML.
• Teacher support: the role of teachers can also be augmented by providing them with AI-based tools to enhance their pedagogical skills, reduce their administrative workload, and facilitate their professional development.
• Curriculum design: AI can help design and update relevant and engaging curricula that reflect the changing needs and demands of the society and the labor market by using data analytics, NLP, and ML.
• Content creation: this is already widely used, as AI can act as an assistant to create and curate high-quality and engaging learning content, such as textbooks, videos, quizzes and games by using NLP, computer vision and DL.
• Career guidance: with its capacity to provide tailor-made suggestions based on large training datasets, AI tools can provide personalized and data-driven career advice and guidance to learners and job seekers by using recommender systems, NLP and ML.

Zero-Knowledge Proofs

Zero-Knowledge Proofs (ZKP) is a cryptographic protocol that enables one party — the prover — to prove to another party — the verifier — that they have some piece of information, without revealing what that information is. In other words, they can prove that they know something without actually telling what it is: “I know a secret, I can prove it, but I won’t say what it is”.

OK that’s the theory, but let’s use a simple example from the popular “Where’s Wally?” game to make that even clearer. Imagine that you and your friend are looking for Wally in a picture. You know where he is, but your friend doesn’t believe you. How can you convince your friend that you know Wally’s location without revealing it to him?

One way to do this is by using a large piece of paper to cover up the entire picture, except for a small cutout showing only Wally. By looking through the cutout, your friend can confirm that Wally is indeed in the picture and that you know where he is, without actually seeing his location relative to the rest of the picture.

That’s called a non-interactive ZKP, where someone can demonstrate knowledge of a particular piece of information without revealing any other information. The person looking through the cutout gains the proof that Wally exists and that you know his location, but does not gain any other knowledge that could be used to locate Wally.

Illustration of zero-knowledge proof concept using “Where’s Wally” example

The concept of ZKP should now be understood, so let’s focus on IRL use cases for which it would bring concrete value. Note that while most of the examples below focus on the privacy-preserving aspects of ZKP, it’s worth highlighting that its usage also provides a lot of benefits on the verifier side in terms of process duration (information accuracy is checked automatically as part of the proof generation) and storage space (the proof weights a fraction of the size of the documents/information it’s proving).

Health & Pharmaceuticals

• Clinical trial data sharing: Using ZKP, researchers and pharmaceutical companies can share and verify the results of clinical trials without revealing sensitive information about the participants, such as their identity, medical history, or genetic data.
• Patient identity verification: On patients’ side, it is possible with ZKP to prove one’s identity to healthcare providers or insurers without revealing personal or medical details, such as name, date of birth, address, or diagnosis.
• Prescription verification: Patients can also prove that they have a valid prescription for a medication without disclosing the name of the medication, the dosage, or the reason for taking it.
• Drug supply chain tracking: As previously with ML, ZKP’s added-value to the supply chain aspect is actually valid for most non-service businesses, not only pharma. Here, manufacturers and distributors could prove that they have followed the proper procedures and regulations for producing and transporting drugs without revealing their trade secrets or confidential information.
• Genomic data analysis: Researchers and biobanks could analyze genomic data from multiple sources without compromising the privacy of the data owners or violating their consent preferences.

Banking & Insurance

• Blacklist verification: ZKP can allow an employer to verify an employee has no financial holdings on a blacklist without revealing the other (allowed) holdings of the employee. This can help prevent conflicts of interest and insider trading while preserving the employee’s privacy.
• Risk constraints verification: A fund can use ZKP to convince its investors that its holdings subscribe to particular risk constraints (VaR, correlation coefficient, various ratios metrics or even carbon footprint or ESG rating), without disclosing the actual holdings. It would increase transparency and trust between the fund and its investors while protecting the fund’s proprietary strategies.
• Aggregate information verification: Once again leveraging on ZKP’s privacy feature, a collection of investors of a fund can verify aggregate information provided by a fund while preserving pairwise anonymity. The purpose here is to reduce fraud and collusion among the fund managers and investors while respecting their individual privacy.
• Creditworthiness/identity verification: A customer can prove their creditworthiness or identity to a bank or an insurer without revealing their personal or financial details. This can help streamline the verification process and reduce the risk of identity theft or data breaches.
• KYC (Know Your Customer): this actually corresponds to adding the previous bullet points all together. KYC is a process by which businesses verify the identity of their customers to comply with anti-money laundering and counter-terrorism financing regulations, the problem here again being that KYC often requires customers to share sensitive personal and financial information with businesses, which can pose privacy and security risks. Using ZKP, a customer could for example prove that they are over 18 years old, that they have a valid bank account, or that they are not on a sanctions list, without disclosing their name, date of birth, account number, or any other information.
• Dark pools: there have been discussions around implementing private trading venues where large investors can trade securities anonymously and without revealing their trading intentions to the public market (private order books), as ZKP can be used to enhance the privacy and efficiency of dark pools by allowing traders to prove that they have a valid order without revealing the details of the order, such as the security, quantity, or price. This could reduce information leakage and market impact, as well as enable faster and cheaper execution of trades.

Education & Professional Training

• Student identity verification: using ZKP, students can prove their identity to online learning platforms or exam proctors without revealing their personal details, such as their name, address, date of birth or gender.
• Credential verification: Students or professionals could prove that they have obtained certain degrees, certificates, or licenses without disclosing the details of their academic records or professional history. Similarly to the first example, the aim here is to dismiss any form of bias.
• Skill assessment: Students or professionals could prove that they have mastered certain skills or competencies without revealing the specific questions or tasks they have completed. It can be very useful for assessments based on real-life situations, real-persons identities or any type of private information (medical secret, national security matters, private intellectual property).
• Privacy-preserving analytics: This is particularly relevant in Europe given its data protection policy — Educators or researchers could analyze student data from multiple sources without compromising the privacy of the students or violating their consent preferences.

Other sectors

• Blockchain (rollups): some specific types of ZKP (notably ZK-SNARKs, i.e. Zero-Knowledge Succinct Non-interactive Argument of Knowledge, and ZK-STARKs, i.e. Zero-Knowledge Scalable Transparent Argument of Knowledge) are already used to enhance the scalability of public blockchains, mainly thanks to their succinctness characteristic. The general concept is known as ZK-Rollup and this is currently a hot topic in the blockchain industry, with some recent deployments in production on mainnet. Here are some of the projects working on the matter: StarkNet, zkSync ERA, Polygon zkEVM, Loopring, Linea, Scroll, Aztec3… They all have various specificities that we won’t cover here, but for more info about rollups (not only ZK ones), here is an article by Vitalik Buterin, co-founder of Ethereum and one of the main thoughts leaders of the crypto-space: https://vitalik.ca/general/2021/01/05/rollup.html.
• Blockchain (privacy): Apart from rollups, ZKP are also a way to enable users to verify transactions without revealing their identities, balances, or inputs and outputs. Such examples of blockchain projects that leverage on this characteristic are Zcash or Monero, which have been around for a few years already. Another interesting product to mention here is Tornado Cash: it is an open source, non-custodial, fully decentralized cryptocurrency mixer that runs on Ethereum (and some compatible networks). It allows users to transfer cryptocurrencies anonymously by using ZKP to hide the origin, destination, and counterparties of their transactions. However, Tornado Cash has been recently blacklisted by the U.S. Treasury for allegedly laundering more than $7 billion in cryptocurrencies, including funds stolen by a hacking group linked to North Korea. This has resulted in the suspension of its domain, GitHub repository, and developers’ accounts, as well as the freezing of some of its funds. However, some pieces of Tornado Cash’ architecture can be reused: going back to the dark pools example from above, a possible implementation would consist in requiring the users to prove that they have a valid identity or membership credential issued by a trusted authority, such as a regulator or a broker, without revealing their actual identity or credential. This would ensure that only authorized users can access the dark pool and that they comply with the relevant rules and regulations, while keeping all the benefits of anonymity and privacy.
• Cybersecurity: ZKP can improve the security and efficiency of authentication protocols, such as passwords, biometrics, or multi-factor authentication, by allowing users to prove their identity without disclosing their credentials or biometric data. Examples of applications leveraging such cybersecurity features enabled by ZKP are Signal or ProtonMail.

Combo: ZKML

Zero-Knowledge Machine Learning (ZKML) is a technique that allows the injection of privacy into ML models. For instance, two parties could train a model together on their combined datasets without sharing the actual data with each other. ZKML is particularly useful in situations where privacy is a concern, as it ensures that sensitive information remains hidden while still enabling collaborative learning. ZKP can indeed enable data owners and analysts to collaborate on data analysis without compromising the privacy or integrity of the data on one side, while data analysts can prove that they have performed certain computations or transformations on the data (typically, in this case, the training of a ML model) without revealing the intermediate or final results. There are actually already a few data analytics platforms that use ZKP along with ML: QEDIT, Inpher, or Duality Technologies.

(Zero Knowledge + Machine Learning) フュージョン = ZKML

But let’s go further and look into some more concrete ideas for potential use cases, focused on our 3 sectors of interest:

Health & Pharmaceuticals

• Collaborative research: ZKP can be used to enable researchers from different institutions to collaborate and train ML models on their combined datasets without sharing the actual data with each other, preserving the privacy of the data while still enabling joint research.
• Clinical trial data analysis: ZKML can allow pharmaceutical companies to analyze clinical trial data while preserving the privacy of patients’ health data.
• Medical device development: In a similar fashion, it can be used by medical device manufacturers to train ML models on sensitive patient data without risking exposure of that data stored in (expensive to maintain) centralized servers..

Banking & Insurance

All the examples below emphasize ZKML’s added-value to collaborative process.

• Risk assessment: the use of ZKML can allow banks and insurance companies to jointly assess risk across their datasets without sharing sensitive customer data.
• Fraud detection: ZKML makes it possible to enable collaboration between banks and other financial institutions to detect fraudulent activity without sharing confidential customer information.
• Customer data analysis: it can also be used to enable banks and insurance companies to jointly analyze customer data to identify patterns and trends without sharing any personally identifiable information — ensuring a potential standardized industry-wide ML training protocol.

Education & Professional Training

• Skills assessment: ZKML can be used to enable employers to assess the skills and knowledge of their employees without exposing any sensitive data.
• Joint training: ZKML can be used to enable multiple organizations to collaborate and train machine learning models on their combined datasets while preserving the privacy of the data.
• Performance evaluation: ZKML can be used to enable joint performance evaluations of employees or students without exposing any personally identifiable information.

Closing Thoughts

In conclusion, all of these technologies, whether ZKP, AI or their intersection ZKML, have significant potential for disrupting many industries, not only the ones covered in the article. They indeed all have very generic characteristics: ZKP offer a way to preserve privacy while still providing proof-of-knowledge, AI can provide insights and automation that were previously unimaginable, while ZKML allows exciting possibilities in terms of secrets-preserving collaboration. The use cases we’ve listed here are just the tip of the iceberg, and we’re excited to see what other innovative applications will emerge as these technologies continue to evolve. What a time to be a technologist!