Building Privacy into KIN, and Personal AI

Part 4 of 4 in the series "The Road to Privacy in Personal AI".

• Part 1: Privacy in Personal AI

• Part 2: Three Core Challenges in Data Encryption

• Part 3: Solving the holy grail of “Encryption in Use”

• Part 4: Building privacy into KIN, and personal AI

This article is about how we decide to do

1. data computation, and

2. data storage

while ensuring privacy and user-controlled data.

In the previous article, we've outlined five different approaches, with the choice of technology depending on the specific use case and requirements.

Our use case focuses on building a trustworthy, personalized AI that remembers your conversations, learns from them, keeps them secret, and is so powerful that you will use it to assist you in life as your companion, assistant, coach, or friend.

3 Requirements that guide us

Let’s look at the 3 main requirements affecting our decision-making direction:

1. KIN must be powerful, fast and cost-effective

2. KIN must work around the clock, even with the app closed

3. KIN must be able to use state-of-the-art Machine Learning Models

1. Power, Speed, and Cost-Effectiveness for KIN

This means leveraging local, on-device computation. Computing data in a cloud environment is very costly, and network latency by using cloud computing can negatively impact speed.

Fortunately, your smartphone is an incredibly powerful machine. Taking advantage of this, and doing as much as possible locally, not only provides a more performant and generally better user experience, but it’s also preferable from a privacy perspective and saves cloud costs.

We’ve seen Apple invest significantly in “Edge Machine Learning” (Edge ML) with their Apple Silicon and Neural Engine, similar to Google/Android and browser APIs such as WebGPU, which give direct access to local GPU. 

2. KIN Must Work Around the Clock

Doing everything locally is great, but sometimes you want actions to occur without needing to keep the app open. This could be asking your AI agent to perform tasks and report back later, or running periodic data imports. Without this capability, we would severely limit our capabilities.

Thus, KIN needs to support background/asynchronous tasks and data processing.

3. KIN requires state-of-the-art ML models

While Edge ML is becoming increasingly powerful, it’s a fact that running the LLMs needed for a top-tier AI agent is currently not possible on edge devices, especially not on mobile. 

KIN must, therefore, have a secure, privacy-preserving way to run models we can’t run on your device.

Deciding on Data Computation

Based on our requirements and computation needs, we can shortlist two candidates for data computation: 

1. Local/On-Device computation

2. Confidential Cloud computation

Thus we have decided on a hybrid architecture that combines local-first with confidential cloud computing.

We utilize local resources for data and computing when possible, offloading to the confidential cloud (Trusted Execution Environments, or TEEs) when necessary, e.g., for LLM inference or async long-running tasks.

Additionally, we're closely monitoring advancements in Fully Homomorphic Encryption (FHE), a technology we intend to explore and utilize once it becomes practical. Integrating FHE with LLM inference is roughly five years away, and we're aligning our experimentation timelines accordingly to harness its full potential when ready.

Data Storage and Privacy

The strategy we choose for processing data (data computation) directly influences the necessary data storage location or its accessibility. Given our three key requirements, the approach is as follows:

1. For Edge ML →  Your data must be available locally, so stored on your device

2. For Cloud ML → Your data should be stored in the cloud

Because we want KIN to be able to do work “in the background” (requirement 2), the data needs to be stored somewhere that can be accessed from the cloud, eg. also in the cloud.

Storing data in the cloud requires two additional challenges to be solved:

1. Maintaining privacy when synchronizing data between your local storage and cloud storage.

2. Providing secure data access to your AI Agent in the cloud without compromising data integrity and privacy.

6 Key Components for storing user data

In developing KIN, a hybrid data storage strategy has been adopted, balancing local-first storage with cloud capabilities to ensure privacy, efficiency, and functionality. Here's a concise overview of the key components:

1. Local-first Data

KIN prioritizes storing data on the user's device, enhancing privacy and reducing latency for faster, local processing. This approach aligns with our commitment to data sovereignty and performance.

2. Synchronization

Through a master-to-master synchronization with a server coordinator, KIN ensures data consistency across devices/clients, facilitating incremental updates and secure backups/snapshots. This allows KIN to offer multi-device access seamlessly and allows a KIN agent running for example in a confidential cloud to sync the data as well.

4. End-to-End Encryption (E2E)

All data transmitted to and stored in the cloud is protected with end-to-end encryption, safeguarding user privacy by making data accessible only to the user, even when leveraging cloud functionalities.

5. Advanced Data Structures

Employing advanced data structures, such as vector embeddings, KIN optimizes data handling for AI processes, supporting sophisticated functionalities like semantic search and personalization at scale.

6. Permission Layer

A robust permission system with multiple keys or clients allows users to control access to their data, ensuring that sensitive information is accessible only to authorized entities and processes within KIN.

This streamlined approach ensures that KIN is not only a powerful AI companion but also a guardian of user privacy and data security, leveraging the best of local and cloud technologies.

Your Data is Your Business, Not Ours

User-controlled data is a key aspect we haven't emphasized enough. KIN empowers you, the user, by ensuring your data is controlled and owned by you, not by KIN. You control the access to your data and you give the permissions to your data.

We don’t want to lock your data in, you should always be able to access it and use it as you please. 

Your data is your business, not ours.

KIN is more than a personal AI

KIN is not just a technological advancement, but a step towards a more secure and private digital future and a future with personal AIs. 

If you value your privacy and wish to take control of your data, we invite you to join the KIN and spread the word.

By getting a KIN, using it, and inviting others on board as well, you'll be part of a movement prioritizing security, privacy, and user empowerment in the digital age.

Go to KIN, your personal ai for your private life