Hey Yoyo! Thanks for this proposal, Manifund is happy to support donations to this project as part of our science & tech research portfolio. Unfortunately, it doesn't seem like a good fit for EA Community Choice, as it's not aimed at helping members in the EA community; I've removed it from consideration for the the quadratic funding match. Best of luck with your work!
1
cryptography for neural data
Project summary
Problem: There is little active research on technically securing neuroimaging data in both the short and long term. Lack of neural security has implications: For example, multiple institutes could have cohorts of Alzheimer’s disease patients and MRI data, however, due to lack of encryption, researchers refrain from sharing data and only work with subsets of public data or from their own labs. The downstream effect is low sample size, a lack of significant conclusions, and disorders that aren’t understood and can’t be diagnosed early.
We need usable, efficient and accurate encryption techniques for each stage of the neural data collection pipeline.
We have validated the lack of security and automation regarding neuroimaging data with 10+ researchers within neuroscience and with neurotech startups as of Aug 2024. We are currently discussing ideas with g.tec, EON systems and the neuroscience director of NexStim.
What are this project's goals? How will you achieve them?
Overview: Select algorithms from research, implement, benchmark and integrate the most appropriate ones into standard neuroimaging software.
First, we need to create a proof-of-concept: We use homomorphic encryption libraries on a sample dataset. The library Microsoft SEAL has HE implementations along with several datasets on openneuro.org. We can simulate the sharing of the encrypted data, simple analyses and visualizing results.
We then need to create a software; Test other homomorphic encryption/secure multiparty computation techniques aiming to select for speed and accuracy. To improve usability, we plan to create a GUI to integrate with existing g.tec software.
Next, we iterate in a real-world setting: Getting data from humans in real-time using the unicorn headset. We can also gather feedback from neuroscience labs, and cryptography experts to continuously iterate.
POC should be completed by Sept 1st, 2024. In order to submit to competition. https://exanova.notion.site/Neurocryptography-share-92c5f00de12948678cb96c968b2f5329?pvs=4
How will this funding be used?
$25k will enable us to work on this project part-time (12 hours a week) for three months since we're both university students. Some of the money will also go toward a BCI headset and software (unicorn headset and g.suite 2020)
Who is on your team? What's your track record on similar projects?
Yoyo Yuan:
enrolled in a quantum computing course and quantum physics seminar in high school, which involve a lot of applied mathematics, in 2021
built introductory projects within brain-computer interfaces in 2021
built a neural network which controls swarms based on haptic hand gestures at UWaterloo in 2022
collaborated with a startup at Founders Inc. on swarm drones and created a simulation in Unity in 2024
have participated in neuroscience research projects through Neuromatch Academy - Computational Neuroscience and NeuroAI courses in 2023, 2024.
Amina Rakhimbergenova:
won the NASA Space Apps Challenge (2023), Treehacks (2024) and the World Robot Olympiad (2021).
cofounded FunCode after graduating high school. FunCode allows students to learn programming concepts and rewards them with crypto. It was recognized as the top 10% of a business accelerator program and was listed among the top 100 Kazakhstani startups.
https://www.linkedin.com/in/aminarhe/
https://www.linkedin.com/in/yoyo-yuan/, https://twitter.com/indiraschka
What are the most likely causes and outcomes if this project fails?
Outcomes
Noisy neural data: Some encryption schemes trade off precision for efficiency (e.g. CKKS). So performing a bunch of operations could make the neural data noisy.
Latency: Cryptographic operations are computationally intensive. There could be time delays in the system, or drain battery life quickly.
I talk about the project in public, so even if failed, it will be a net positive in introducing more people to neurosecurity, an underdeveloped field. If this project doesn't gain enough traction I will be likely working on cryptography for an adjacent data-heavy field and transferring it to neuroimaging data.
What other funding are you or your project getting?
none, I wanted to gain momentum on the project first
For any questions, please contact @indiraschka on Twitter.
🥦
Tony Gao
3 months ago
My understanding is that HE is extremely slow as a result highly constrained in the operations that are supported. What are the simple analyses which are allowed after the imaging data is encrypted? If they are not comprehensive, is it possible that if your project succeeds and the highest quality open data set is encrypted, that the default suite of analysis will be far more limited and we will potentially miss out on discoveries which could be uncovered by more sophisticated analytical techniques?
wasabipesto
3 months ago
I am unfamiliar with both fields involved here (neuroimaging and cryptography) so I had to do some digging to understand this proposal. Here’s my understanding of the situation and some basic questions, let me know if I’m misunderstanding:
The problem is that researchers aren’t sharing anonymized brain imaging data. This is already legal but rarely done in practice due to privacy concerns.
What types of advancements are lacking good brain scan data? What sorts of research is this data good for?
The solution is to apply a special form of encryption to the brain scan data. When the data is encrypted this way it allows researchers to do some stuff with the data but not other stuff.
What exactly can you do with an encrypted scan vs a non-encrypted scan? If it’s already anonymized, what danger exists in a non-encrypted scan that’s removed when it’s encrypted?
There are a lot of ways to do this special encryption. The first part of this project is to take some existing brain scans and try each way of encrypting them, hopefully finding one that works but is still easy to use for researchers.
What is your list of HE implementations? What criteria do they need to pass in order to be useful? What sort of performance impact do you expect from this intermediate processing step?
The last part of this project is to take a live brain scan and encrypt it in real-time so it’s never unencrypted.
What benefit does this have? Shouldn’t original researcher should still have full access to the original scan?