Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EncGPT: A Multi-Agent Workflow for Dynamic Encryption Algorithms

Donghe Li, Zuchen Li, Ye Yang, Li Sun, Dou An, Qingyu Yang

TL;DR

EncGPT introduces a dynamic, multi-agent framework that uses LLM-driven rule, encryption, and decryption agents to generate and apply encryption rules within a decentralized dialogue workflow, aiming to balance security with computational cost. By leveraging classical algorithms for rule generation and a three-round prompting strategy, the system tests homomorphic properties and evaluates GPT-4o's performance in rule creation and encryption/decryption tasks. The results demonstrate feasibility with substitution ciphers and highlight challenges such as model bias, hallucinations, and limited arithmetic capabilities, suggesting further optimization and tool integration. This work advances LLM-MA research by showing how dynamic encryption rules can be produced on-the-fly, potentially enhancing security in adaptive communication scenarios while outlining clear directions for future improvements and broader cryptographic coverage.

Abstract

Communication encryption is crucial in computer technology, but existing algorithms struggle with balancing cost and security. We propose EncGPT, a multi-agent framework using large language models (LLM). It includes rule, encryption, and decryption agents that generate encryption rules and apply them dynamically. This approach addresses gaps in LLM-based multi-agent systems for communication security. We tested GPT-4o's rule generation and implemented a substitution encryption workflow with homomorphism preservation, achieving an average execution time of 15.99 seconds.

EncGPT: A Multi-Agent Workflow for Dynamic Encryption Algorithms

TL;DR

EncGPT introduces a dynamic, multi-agent framework that uses LLM-driven rule, encryption, and decryption agents to generate and apply encryption rules within a decentralized dialogue workflow, aiming to balance security with computational cost. By leveraging classical algorithms for rule generation and a three-round prompting strategy, the system tests homomorphic properties and evaluates GPT-4o's performance in rule creation and encryption/decryption tasks. The results demonstrate feasibility with substitution ciphers and highlight challenges such as model bias, hallucinations, and limited arithmetic capabilities, suggesting further optimization and tool integration. This work advances LLM-MA research by showing how dynamic encryption rules can be produced on-the-fly, potentially enhancing security in adaptive communication scenarios while outlining clear directions for future improvements and broader cryptographic coverage.

Abstract

Communication encryption is crucial in computer technology, but existing algorithms struggle with balancing cost and security. We propose EncGPT, a multi-agent framework using large language models (LLM). It includes rule, encryption, and decryption agents that generate encryption rules and apply them dynamically. This approach addresses gaps in LLM-based multi-agent systems for communication security. We tested GPT-4o's rule generation and implemented a substitution encryption workflow with homomorphism preservation, achieving an average execution time of 15.99 seconds.

Paper Structure

This paper contains 16 sections, 6 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Workflow Architecture
  3. Encryption Rules
  4. Encryption and Decryption
  5. Homomorphism
  6. Experiment
  7. Rule Preferences
  8. Encryption and Decryption
  9. Conclusion
  10. Limitations
  11. Prompt
  12. Rule Agent
  13. Encryption Agent
  14. Decryption Agent
  15. Agent Communication
  16. ...and 1 more sections

Figures (6)

  • Figure 1: The architecture diagram of our workflow is shown. Green lines and the matrix represent the encrypted communication flow. On the left, the Recipient Agent interacts with the external environment, while on the right, agents communicate with each other. Red lines indicate the transmission of ciphertext, and blue lines indicate the transmission of plaintext.
  • Figure 2: GPT-4o’s Preferences in Encryption Rule Generation
  • Figure 3: Dialogue Used for Rule Generation
  • Figure 4: Dialogue Used for Encryption
  • Figure 5: Dialogue Used for Decryption
  • ...and 1 more figures