01
LLM Capsule is a security solution
that allows users to use Public LLMs such as ChatGPT
without worrying about exposing their personal and sensitive information.
LLM CAPSULE
- Personal Info / Credit Info / Sensitive Info / Confidential…
- Resolve all issues at once!
02
LLM Capsule automatically detects personal and sensitive info
contained in text and encapsulates it securely.
03
LLM Capsule allows anyone can experience
the stability of a Private LLM and the superior performance of a Public LLM
at a low cost.
GS Certification
LLM Capsule v1.0 has been certified as GS Grade 1. (24-0308)
It is sold at the Public Procurement Service in Korea.
*GS Certification : Good Software Certification in Korea
PROCESS
LLM Capsule is designed to allow users to process documents securely
and users can receive responses immediately based on converted documents.
< Capsule & Uncapsule Process >
USAGE
Compliance with
Personal Info De-identification Guideline
Personal/Sensitive Info
100% protected
Encapsulation Technology + Prompt EngineeringAccuracy 98% (Benchmark Dataset)
- Maintain LLM response performance even with security processing
- 20% improvement in response accuracy (VS masking-based)
- Availiable at 1/300 price (VS Private LLM)
- The guideline is the standards of 'Ministry of the Interior and Safety' of Korea.
01 Original Document
(Containing sensitive info)
Name: Smith / Age: 29 (Joined: Mar 1, 2023)
Major: Computer Science / Dept.: IT
E-mail: smith@cubig.ai
Name: Johnson / Age: 38 (Joined: Sep 22, 2022)
Major: Economics / Dept.: Planning
E-mail: johnson@cubig.ai
Name: Brown / Age: 27세 (Joined: Jan 2, 2024)
Major: Mathematics / Dept.: IT
E-mail: brown@cubig.ai
...02 Safety Transformed
( Capsuled sensitive info)
03 Input Prompt
(LLM Capsule + Public LLM)
04 LLM Result
(Protected sensitive info)
05 Restored Result
(Contained sensitive info)
Privacy performance verificaton
Detection Type 1 errors : 98.10%(Accuracy)* Verification of 6 types of personal info recognition technology
Detection Type 2 errors : 99.14%(Accuracy)
※ Detection and protection of formalized personal info : 100%
* Verification of detected and encapsulated
* Error types are divided into 1 and 2
| Errors | Description |
|---|---|
| Type 1 |
Detection failed (Test with 105 sentences containing sensitive info) |
| Type 2 |
Incorrect encapsulation (Test with 1,402 sentences that do not contain sensitive info) |
LLM Utility performance verification
Similarity of LLM Capsule & Public LLM : About 98%* Send encapsulated data instead of original data to LLM to query
※ Privacy-Utility Trade-off is improved
* LLM Utility impact test (+masked data)
* Comparison of LLM Capsule and Public LLM response
* Test with 80 documents/questions (20 documents/4 questions/total 80 datasets)
* The original text is fake personal info.
| Test | Settings |
|---|---|
| ChatGPT | Save the responses : When sending original text with ChatGPT |
| LLM Capsule | Save the responses : When a sentence encapsulated(anonymized) with LLM Capsule is sent to ChatGPT 3.5 |
| Comparison of similarity | Comparison of ChatGPT and LLM Capsule response similarity with LLM |
PC Requirements
Stable operation even with minimum specifications
- Windows 10 or later
- RAM 4G or more
FEATURES
Existing LLM compatible
You can use your existing LLM as is
Cost savings
Available at 1/300 price compared to Private LLM
Protect confidential
Encapsulate and protect company's sensitive info
Intranet compatible
Available for use on company's intranet
