This is the age of AI, and I am here to do another examination of decentralized AI (DeAI). Today, in this 2-part series, I will look into the way trusted execution environments (TEEs) can power verifiable privacy. For my inputs, I will use references from Oasis Labs and Oasis Protocol.
Oasis has been developing AI frameworks as an extension of its privacy-first philosophy and expertise long before the cryptoAI landscape gained prominence. In an earlier article, I had briefly discussed adopting the decentralized approach in working with LLMs as a potential game-changer over the traditional, centralized setup.
One of the most critical advantages of DeAI is the ability to provide provenance for the AI models. As a result, we gain valuable insights that espouse the blockchain principles of transparency and immutability:
- Source, as in which pre-trained foundation model is used
- Method, as in what additional training steps are used to specialize the model
- Content, as in what training data has been used
Oasis posits that GPU-enabled trusted execution environments (TEEs), combined with its self-developed runtime off-chain logic (ROFL) framework, can help build sustainable, specialized AI models with verifiable provenance information published on-chain. While the runtime on-chain logic is handled by Oasis Sapphire's utilization of TEEs powered by Intel SGX v1, Oasis ROFL leverages confidential computing support with TDX-powered TEEs that can perform GPU-accelerated ML training and inference tasks in an integrity-protected, attestable environment.
This clears the path to introduce decentralized marketplaces for AI models and services, with built-in openness, transparency, and community governance, and USPs such as confidentiality of sensitive data, freedom from bias, and fair compensation for use of data and models.
In the concluding part of the series, we will illustrate a proof of concept by demonstrating step by step how to use off-chain GPU-enabled TEEs and ROFL to create specialized AI models.
Resources to explore further:
Top comments (0)