Homoglyph-based method to prevent AI cheating in CS homework
A new paper on arXiv proposes using homoglyph-based adversarial perturbation to prevent students from using AI tools like ChatGPT, Gemini, and Claude to solve introductory computer science theory problems. The method modifies questions by replacing characters with visually similar homoglyphs without changing the semantic meaning, making it harder for AI to produce correct answers while remaining readable to humans. The researchers developed an interactive tool to apply this technique. The paper is categorized under Computer Science > Computers and Society and was submitted to arXiv.
Key facts
- Paper proposes homoglyph-based adversarial perturbation to prevent AI cheating in CS homework.
- Method modifies questions by replacing characters with homoglyphs without changing semantic meaning.
- Targets AI tools like ChatGPT, Gemini, and Claude.
- Focuses on introductory computer science theory problems.
- Researchers developed an interactive tool for the method.
- Paper is on arXiv under Computer Science > Computers and Society.
- Submission history available on arXiv.
- Experimental results show theoretical problems can be effectively perturbed.
Entities
Institutions
- arXiv