Don’t Trust It Until You Own It: Trusted Compute
With the rise of self-sovereign identity, there’s a shift in data ownership. It’s now about owning your data, finding services that host it, and ensuring the framework solves this problem for you. But what about the derivatives of your data? What about computing? How do you ensure the process is secure when you need to use your data?
Encryption vs. Homomorphic Encryption
Encryption:
- Purpose: Primarily used to protect data at rest or in transit.
- Functionality: Once data is encrypted, it remains locked until decrypted with the appropriate key.
- Limitation: To perform any operation on the encrypted data, it must first be decrypted, which can expose the data to potential threats.
Homomorphic Encryption:
- Purpose: Designed to allow computations on encrypted data without decrypting it first.
- Functionality: Enables operations directly on the encrypted data, producing an encrypted result that, when decrypted, matches the result of the functions as if they had been performed on the plaintext.
- Advantage: Offers higher data security, especially in cloud environments, as the data remains encrypted even during computation.
The challenge with using just encrypted data for computations is that traditional encryption methods require decryption before any operation. This decryption step can expose sensitive data. Homomorphic encryption overcomes this by allowing computations on encrypted data, ensuring data remains secure throughout the process.
Homomorphic Encryption in Detail
Homomorphic encryption is a fascinating topic in computation, especially when combined with AI. A notable example of its application with AI can be found
The challenge with homomorphic encryption, especially when combined with machine learning, is that you have a limited number of operations. For instance, in one schema, you might perform homomorphic addition, while in another, you can do homomorphic multiplication. Combining these operations is challenging and computationally heavy. Currently, the precision of models using homomorphic encryption is around 70%, and they can’t solve highly complex problems.
Zero Knowledge Proof Machine Learning (ZKML)
Another intriguing domain is the zero-knowledge proof machine learning (ZKML). It’s another method being explored, but it’s still computationally heavy, complex, and not yet ready for production.
Lear more
Trusted Execution Environment
Then there’s the domain of confidential computing, confidential containers, and trusted execution environments. Nowadays, almost everything is hosted in containers. But how can you ensure that the memory and CPU areas are not transparent to the host machine? How can you be sure that big providers like Amazon don’t have access to your data?
Trusted execution environments provide an answer to this. However, the challenge is that these environments often depend on specific hardware, which can be expensive or limit your choices in cloud providers. Especially for machine learning, where you need not just a CPU but also a GPU, building a trusted execution environment with protected GPU hardware becomes essential.
Conclusion
The topic of trusted compute is vast, and we’re just beginning this journey. It’s crucial to build trusted protocols and trusted compute together. As the saying goes, “Don’t trust it if you don’t own it.” The future lies in more secure and owner-centric compute.
— -
