Semantic Cipher

Encode arbitrary data into semantic text

Plaintext: 0xdeadbeef
Encoded text: Launching every day, fearless astronauts will always conquer all universe. Astronauts wearing advanced technology aboard craft achieve cosmic adventures always

Index

1. How it Works
   1.1. Encryption
         1.1.1. Step 1. Convert Plaintext to Hexadecimal Format
         1.1.2. Step 2. Create Hexadecimal Mapping
         1.1.3. Step 3. Map Each Corresponding Hexadecimal Value to English Word
   1.2. Decryption
         1.2.1. Step 1. Parse the Ciphertext and Extract First Character of Each Word
         1.2.2. Step 2. Map the Hexadecimal Characters to Their Corresponding Values
2. Examples
   2.1. Encrypt Using OpenAI Models
   2.2. Encrypt Using Pretrained SLM

How it Works

Encryption

Step 1. Convert Plaintex to Hexadecimal Format

• The plaintext is converted into a hexadecimal representation

  • Input: Hello

    • H → ASCII: 72 → Hex: 0x48
    • e → ASCII: 101 → Hex: 0x65
    • l → ASCII: 108 → Hex: 0x6C
    • l → ASCII: 108 → Hex: 0x6C
    • o → ASCII: 111 → Hex: 0x6F

    Output: 0x48656C6C6F

Step 2. Create Hexadecimal Mapping

• Each of the 16 hexadecimal characters are mapped to one of the 16 highest frequency English characters. These characters are E, T, A, O, I, N, S. H, R, D, L, C, U, M, W and F.

• If a key is a chosen the hex mapping will be "scrambled" in such a way that only the key can be used to restructure the hex mapping to retrieve the semantically encrypted text.

• Hex mapping when key=""

  Hex	Character
  0	U
  1	H
  2	T
  3	I
  4	L
  5	R
  6	A
  7	F
  8	E
  9	N
  A	M
  B	C
  C	S
  D	O
  E	D
  F	W

• Hex mapping when key="abc"

  Hex	Character
  0	L
  1	T
  2	H
  3	D
  4	S
  5	F
  6	R
  7	O
  8	E
  9	I
  A	A
  B	M
  C	U
  D	N
  E	C
  F	W

Step 3. Map Each Corresponding Hexadecimal Value to English Word

• For each hexadecimal character in the hexadecimal representaion of Hello (48656C6C6F) lookup the corresponding value and semantically generate the next most probabilistic word. The LLM/SLM will be "forced" to output the next most logical word given the prior words that start with the corresponding character.

  • Input: 48656C6C6F andKey=""
    • 0x4 → Value: L → Learning
    • 0x8 → Value: E → every
    • 0x6 → Value: A → aspect
    • 0x5 → Value: R → requires
    • 0x6 → Value: A → a
    • 0xC → Value: S → strong
    • 0x6 → Value: A → analytical
    • 0xC → Value: S → skill
    • 0x6 → Value: A → and
    • 0xF → Value: W → willpower

  Output: Learning every aspect requires a strong analytical skill and willpower

Decryption

Step 1. Parse the Ciphertext and Extract First Character of Each Word

• Input: Learning every aspect requires a strong analytical skill and willpower and Key=""

  • Learning → Value: L → Hex: 0x4
  • every → Value: E → Hex: 0x8
  • aspect → Value: A → Hex: 0x6
  • requires → Value: R → Hex: 0x5
  • a → Value: A → Hex: 0x6
  • strong → Value: S → Hex: 0xC
  • analytical → Value: A → Hex: 0x6
  • skill → Value: S → Hex: 0xC
  • and → Value: A → Hex: 0x6
  • willpower → Value: W → Hex: 0xF

  Output: 0x48656C6C6F

Step 2. Map the Hexadecimal Characters to Their Corresponding Values

• Input: 0x48656C6C6F and Key=""

  • Hex: 0x48 → Character: H
  • Hex: 0x65 → Character: e
  • Hex: 0x6C → Character: l
  • Hex: 0x6C → Character: l
  • Hex: 0x6F → Character: o

  Output: Hello

Examples

Encrypt Using OpenAI Models

To use OpenAI models, simply add your OpenAI API key to the .env file.

The encrypt method requires one parameter, plaintext, which is the textual data that is to be semantically enciphered.

Two optional parameters can be used:

• The context param notifies the LLM that the generated output should be relevant to the given topic.
• The key param shuffles the hex mapping so that the end user must know the key in order for the text to be decrypted.

There are 16! total permutations, given that the encoding list contains all hexadecimal characters.

sc = SemanticCipher(model_name="gpt-4o", key="xyz")

ciphertext = sc.encrypt(plaintext="0xdeadbeef", context="Space")
print(f"Ciphertext: {ciphertext}")

plaintext = sc.decrypt(ciphertext=ciphertext)
print(f"Plaintext: {plaintext}")

Output:

Ciphertext: Launching every day, fearless astronauts will always conquer all universe. Astronauts wearing advanced technology aboard craft achieve cosmic adventures always

Plaintext: 0xdeadbeef

Encrypt Using Pretrained SLM

Important

• Using SLMs typically output text that is nonsensical. Next token prediction is strictly used and does not leverage the reasoning capabilities of larger models to formulate outputs. Added as an experiment and as a template for future experiments.

sc = SemanticCipher(model_name="Qwen/Qwen2.5-1.5B-Instruct", from_pretrained=True, key="xyz")

ciphertext = sc.encrypt(plaintext="0xdeadbeef")
print(f"Ciphertext: {ciphertext}")

plaintext = sc.decrypt(ciphertext=ciphertext)
print(f"Plaintext: {plaintext}")

Output:

Ciphertext: Lily E. Duffin F. A. W. A. C. A. U. A. W. A. T. A. C. A. C. A. A.

Plaintext: 0xdeadbeef

Tip

• Use requirements.txt to install all necessary packages

pip install -r requirements.txt

• Python version 3.10.15 was used in testing