Skip to main content

hackagent.attacks.techniques.static.mml.image_encoder

Image encoding module for the MML attack.

Generates images containing encoded harmful prompts using various visual transformation techniques:

  • Word Replacement: renders text with substituted words
  • Mirror: renders text then flips horizontally
  • Rotate: renders text then rotates 180 degrees
  • Base64: encodes text to Base64 then renders the encoded string

All images are returned as base64-encoded data URLs suitable for inclusion in multimodal LLM API requests.

encode_word_replacement​

def encode_word_replacement(prompt: str,
                            num_replacements: int = 3,
                            image_width: int = 800,
                            image_height: int = 400,
                            font_size: int = 24,
                            background_color: str = "white",
                            text_color: str = "black") -> Dict[str, Any]

Encode a prompt using the word replacement technique.

Replaces key words with random substitutes, renders the modified text to an image, and returns the replacement dictionary.

Arguments:

  • prompt - The original harmful prompt.
  • num_replacements - Number of words to replace.
  • image_width - Width of generated image.
  • image_height - Height of generated image.
  • font_size - Font size for text rendering.
  • background_color - Image background color.
  • text_color - Image text color.

Returns:

Dictionary containing:

  • image_data_url: base64-encoded image data URL
  • replacement_dict: mapping of replacement → original words
  • scrambled_words: shuffled list of original words
  • modified_text: the text as rendered in the image

encode_mirror​

def encode_mirror(prompt: str,
                  image_width: int = 800,
                  image_height: int = 400,
                  font_size: int = 24,
                  background_color: str = "white",
                  text_color: str = "black") -> Dict[str, Any]

Encode a prompt by rendering it in a horizontally mirrored image.

Arguments:

  • prompt - The original harmful prompt.
  • image_width - Width of generated image.
  • image_height - Height of generated image.
  • font_size - Font size for text rendering.
  • background_color - Image background color.
  • text_color - Image text color.

Returns:

Dictionary containing:

  • image_data_url: base64-encoded mirrored image data URL
  • scrambled_words: shuffled list of original words

encode_rotate​

def encode_rotate(prompt: str,
                  image_width: int = 800,
                  image_height: int = 400,
                  font_size: int = 24,
                  background_color: str = "white",
                  text_color: str = "black") -> Dict[str, Any]

Encode a prompt by rendering it in a 180-degree rotated image.

Arguments:

  • prompt - The original harmful prompt.
  • image_width - Width of generated image.
  • image_height - Height of generated image.
  • font_size - Font size for text rendering.
  • background_color - Image background color.
  • text_color - Image text color.

Returns:

Dictionary containing:

  • image_data_url: base64-encoded rotated image data URL
  • scrambled_words: shuffled list of original words

encode_base64​

def encode_base64(prompt: str,
                  image_width: int = 800,
                  image_height: int = 400,
                  font_size: int = 24,
                  background_color: str = "white",
                  text_color: str = "black") -> Dict[str, Any]

Encode a prompt by Base64-encoding it and rendering the result in an image.

Arguments:

  • prompt - The original harmful prompt.
  • image_width - Width of generated image.
  • image_height - Height of generated image.
  • font_size - Font size for text rendering.
  • background_color - Image background color.
  • text_color - Image text color.

Returns:

Dictionary containing:

  • image_data_url: base64-encoded image data URL (image contains the Base64-encoded text)
  • scrambled_words: shuffled list of original words

encode_mixed​

def encode_mixed(prompt: str,
                 num_replacements: int = 3,
                 image_width: int = 800,
                 image_height: int = 400,
                 font_size: int = 24,
                 background_color: str = "white",
                 text_color: str = "black") -> Dict[str, Any]

Encode a prompt using word replacement, mirror, and rotation combined.

Replaces key words in the prompt with random substitutes, renders the modified text to an image, then mirrors horizontally and rotates 180°. Combines all three visual obfuscation techniques.

Arguments:

  • prompt - The original harmful prompt.
  • num_replacements - Number of words to replace.
  • image_width - Width of generated image.
  • image_height - Height of generated image.
  • font_size - Font size for text rendering.
  • background_color - Image background color.
  • text_color - Image text color.

Returns:

Dictionary containing:

  • image_data_url: base64-encoded mixed-transform image data URL
  • replacement_dict: mapping of replacement → original words
  • scrambled_words: shuffled list of original words
  • modified_text: the text as rendered in the image

encode_prompt​

def encode_prompt(prompt: str, encoding_mode: str,
                  **kwargs: Any) -> Dict[str, Any]

Encode a harmful prompt into an image using the specified mode.

Arguments:

  • prompt - The original harmful prompt text.
  • encoding_mode - One of "word_replacement", "mirror", "rotate", "base64", "mixed".
  • **kwargs - Additional parameters passed to the encoder (image_width, image_height, font_size, etc.).

Returns:

Dictionary with encoding results (always includes "image_data_url").

Raises:

  • ValueError - If encoding_mode is not recognized.