Zero-Knowledge Machine Learning
Organisatsiooni nimi
Cryptography Research Group, ICS
Kokkuvõte
Zero-knowledge (ZK) proofs, especially zk-SNARKs (succinct non-interactive proofs of knowledge), are currently a very trendy topic, especially motivated by the application of privacy-preserving blockchain. Other applications include privacy-preserving machine learning (including the training), privacy-preserving image processing, etc. There is a multibillion-dollar industry. Zk-SNARKs allow one to prove that some large-scale computation was performed correctly with two additional objectives: (1) nothing is leaked about the private data, and (2) verifying the proof should be considerably more efficient than doing the computation itself. The area is currently very rich in theory (with hundreds of research papers annually) and practice (a few hundred active startups). Helger is also currently giving a course (https://courses.cs.ut.ee/2024/zk/fall/Main/HomePage) on ZK proofs.
ZKML is the application of zero-knowledge proofs within machine learning. The idea is to allow someone to prove they ran a machine learning model on some data and got a specific result — without revealing either the data itself or the model’s inner workings. In more technical terms, zkML enables privacy-preserving inference and model verification:
Privacy-preserving inference: You can run a machine learning model
on private data and prove that the model was run correctly without revealing the data or the model’s outputs.
Model verification: You can prove that a model was trained in a
certain way or that its outputs are correct without disclosing the model’s parameters or the dataset used for training.
We expect the student to write a survey on specific ZKML techniques and preferably start independent research. The survey and research should combine ML and ZK knowledge, explaining how to represent common ML tasks (like inference) in the form of amicable to efficient ZK proofs, and explain at least one ZK approach that is amicable to ZKML. Interested students are recommended to browse YouTube videos at https://www.youtube.com/results?search_query=zkml
Funding is covered by a grant, and we hope the student will continue to do a Ph.D.
ZKML is the application of zero-knowledge proofs within machine learning. The idea is to allow someone to prove they ran a machine learning model on some data and got a specific result — without revealing either the data itself or the model’s inner workings. In more technical terms, zkML enables privacy-preserving inference and model verification:
Privacy-preserving inference: You can run a machine learning model
on private data and prove that the model was run correctly without revealing the data or the model’s outputs.
Model verification: You can prove that a model was trained in a
certain way or that its outputs are correct without disclosing the model’s parameters or the dataset used for training.
We expect the student to write a survey on specific ZKML techniques and preferably start independent research. The survey and research should combine ML and ZK knowledge, explaining how to represent common ML tasks (like inference) in the form of amicable to efficient ZK proofs, and explain at least one ZK approach that is amicable to ZKML. Interested students are recommended to browse YouTube videos at https://www.youtube.com/results?search_query=zkml
Funding is covered by a grant, and we hope the student will continue to do a Ph.D.
Lõputöö kaitsmise aasta
2024-2025
Juhendaja
Helger Lipmaa
Suhtlemiskeel(ed)
eesti keel, inglise keel
Nõuded kandideerijale
It's a topic for an MSc thesis. Students will need to have some background in cryptography
Tase
Magister
Märksõnad
Kandideerimise kontakt
Nimi
Helger Lipmaa
Tel
E-mail
Vaata lähemalt