Replacement Text Using Homoglyphs to Disrupt AI Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained on unauthorized content are difficult to detect and remove unauthorized influences, as the output does not directly recreate training data, making it challenging to prevent copyright violations and unauthorized use.
Innovation Solution
A system and method that generates replacement text by replacing character portions of initial keywords with visually similar but machine-distinguishable characters, such as homoglyphs and zero-width elements, to create replacement keywords that obstruct parsing and training by machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models are trained on content from the Internet without permission, then the models can learn from large amounts of data, but this leads to copyright violations and unauthorized use of content
Solution Approach 1:
The system applies preliminary anti-action by proactively replacing characters in the training data before the model processes it. Homoglyphs and zero-width characters are substituted in advance to prevent the model from learning unauthorized content patterns, thereby preventing copyright violations before they can occur during model training or inference
2Difficulty of detecting and measuring
If output of machine learning models is monitored to identify unauthorized training, then copyright violations can be detected, but this is difficult to implement because models do not directly recreate training data
Solution Approach 1:
The system performs preliminary action by embedding detectable markers (homoglyphs and zero-width characters) into the training data before the model processes it. These markers are designed to be imperceptible to humans but detectable by systems, enabling future detection of unauthorized training without requiring complex analysis of model outputs
Solution Approach 2:
The system uses homoglyphs and zero-width characters as intermediaries between the original content and the model training process. These intermediary characters serve dual purposes: they allow the model to process the text normally while simultaneously providing detectable signatures that can trace the content back to its unauthorized source
3Object-affected harmful factors
If replacement text uses visually similar characters to obstruct machine parsing, then unauthorized training can be prevented, but this may affect human readability and accessibility
Solution Approach 1:
The system applies local quality by selectively replacing only specific characters within the training data while leaving the overall text structure and most characters unchanged. This localized replacement of specific character portions with homoglyphs maintains human readability while sufficiently disrupting machine parsing to prevent unauthorized training
Data Source
AI summary
System and methods for generating replacement content are disclosed. Replacement text is generated by replacing at least one respective character portion of a first instance of an initial keyword with a first set of replacement characters to generate a first replacement keyword. At least one respective character portion of a second instance of the initial keyword is replaced with a second set of replacement characters to generate a second replacement keyword. Machine encodings of the first replacement keyword, second replacement keyword, and the initial keyword are distinct. In response to receiving a request for initial text, instructions are generated to display, via a human readable user interface, replacement text including the first replacement keyword and the second replacement keyword.