Cover photo

Ritual Revealed: The Truth About AI Privacy and Data Control

Your Data, Their Rules.

Many of us have started to wonder whether it’s worth giving our data to giants like OpenAI or Google. Their closed models are a kind of “black box”: we don’t see how exactly the algorithms work, what data is collected, and how it is used afterward. Control and transparency are minimal, while tracking of our actions remains entirely under their control.

On the other hand, the emergence of open models like DeepSeek, Mistral, and others has provided a real alternative. Now it’s possible to run LLMs without relying on big corporations and without directly handing over our data to them.

But there is a serious catch these models require huge computational resources.

In the end, freedom exists but it is available only to those who have massive data centers and budgets. Freedom for ordinary users remains out of reach for now.

The Privacy Gap in Open Models.

Today, it’s possible to run LLMs without OpenAI. Thanks to open models like DeepSeek R1, Mistral, or LLaMA 3, we have a real alternative to centralized APIs from OpenAI, Anthropic, and Google. This allows us to avoid dependence on centralized services and maintain control over our own data.

But this naturally raises a question:

Is it possible to run these models privately without the risk of data leaks?

Ensuring privacy when using such models remains a crucial issue.

In the Web3 community, people often talk about “private AI,” but in practice, it’s not that simple there are many technical challenges, especially when it comes to privacy-preserving methods like SMPC. These methods are reliable but very slow and expensive.

And here comes Ritual. They claim that private AI is possible. They are working on innovations that will make private AI more efficient and accessible.

SMPC: Privacy at a Cost.

SMPC, or Secure Multi-Party Computation, is a method where a data processing task is split among multiple servers. Each server only receives a fragment of the information, so none of them sees the complete original input.

In other words, when you run an AI model using SMPC, your data is never fully exposed to any single server. The computation is done collaboratively, while privacy is preserved.

It sounds impressive and it seems like privacy has finally taken on a real, tangible form. However, in practice, things are much more complicated. These technologies are extremely complex and face serious challenges in terms of performance and speed. Computations using SMPC are significantly slower than standard inference, especially when it comes to complex nonlinear operations in large language models.

This means that while SMPC provides a strong level of privacy, it’s still difficult to use it effectively for truly large and powerful models and this is exactly where there’s room for further research and innovation.

SMPC: Challenges and Limitations.

The main problem of SMPC is the nonlinear functions in large language models.Operations like GELU and softmax are poorly "digested" by cryptography.They can't just be computed in parts without a strong slowdown and communication overload.

To solve this problem, researchers started replacing nonlinearities with approximations.

MPCFormer was the first to propose replacing them with quadratic functions.This gave a 2–5× speed increase, but the model loses accuracy and to bring it back, you need to retrain the model, which is expensive and long.

PUMA went further: it uses piecewise polynomial approximations, where each nonlinearity is replaced with four segments of simple formulas.This allows avoiding retraining and achieving high accuracy.But some of these formulas are of the sixth degree, and that is still too resource-intensive in SMPC.

NIMBUS is the last in the chain.It analyzes the distribution of inputs and picks lighter approximations, lowering the degree of polynomials to two.This gives a serious speed boost, while keeping the accuracy, and makes the approach more practical.

Progress is obvious from MPCFormer to NIMBUS, inference was accelerated almost 10 times.But even with such improvements, it is still much slower than the standard model run.The nature of the task is still not on our side: SMPC computations by definition require many steps and interaction between participants, and it’s still impossible to completely avoid that.

Shuffling Privacy: The Power and Limits of Permutations.

Against the backdrop of the complexity and high cost of cryptographic methods such as SMPC, alternative approaches are emerging that aim to ensure privacy in simpler ways.

One such approach is permutation methods. The idea is not to encrypt data or distribute it using expensive protocols, but simply to shuffle the model’s hidden states before sending them for inference. This is possible thanks to the permutation equivariance of LLMs: large language models are robust to changes in the order of internal variables and can correctly process shuffled hidden states. Thus, the model can handle a “shuffled” state, while the server does not receive information about the real data.

The main argument in favor of the security of this approach is the astronomical size of the permutation space. If a hidden state contains 1024 elements, the number of possible permutations is 1024! (factorial of 1024), which exceeds the number of atoms in the Universe. Such a huge number of possibilities makes brute-forcing all permutations and restoring the original order practically impossible.

However, Ritual claims that this is not enough. Even with such a huge space of possibilities, they found a way to recover the original hidden states with almost 100% accuracy and in linear time. They promise to reveal the details later and this calls into question the entire line of research based solely on permutations as a replacement for cryptography.

Ritual’s Attack: Shuffling Isn’t Enough.

Despite the apparent effectiveness of permutation methods, the Ritual team conducted research and showed that they are not as reliable as previously thought. They devised an attack that almost completely recovers the original hidden states with 99.99% accuracy.

This means that even if the data was shuffled before processing, an attacker can still almost precisely determine what the original input was.

Ritual demonstrated that simple data shuffling does not provide reliable privacy and calls into question the effectiveness of all such approaches.

Choosing Between Privacy and Performance.

Right now, the situation is this: if you want truly private AI, get ready for slow and expensive SMPC operations. If you want speed, you have to compromise on security and use methods like permutations which Ritual has already shown don’t provide sufficient protection.

In the end, we face a dilemma: either wait for someone to make a real breakthrough, or choose between usability and true privacy.

We’re in a situation where you either sacrifice privacy for speed, or wait for new technologies that can provide both.

We’re one step away from a breakthrough one where you no longer have to choose between privacy and convenience. I’m sure Ritual will eventually show how to make it happen.

Ritual’s Promise: A New Era of Private AI.

After demonstrating the attack on permutation-based methods, the Ritual team announced that they are working on their own solution to the problem of private inference.The promises are ambitious:

  • Faster than SMPC

  • More reliable than simply shuffling hidden states

  • At the same time, open and verifiable

The technical details have not yet been disclosed, but the focus is on accessibility, verifiability, and practicality which distinguishes their approach from previous complex or theoretical solutions.If they succeed in implementing this in practice, it could lead to a real breakthrough in AI privacy — not just in theory, but in ready-to-use tools.