Do LLMsMakeNeural Distinguishers Wise?
Researchers explored using large language models (LLMs) to enhance neural distinguishers, a cryptanalysis technique for symmetric-key cryptography. Their experiments on the SPECK-32/64 cipher revealed that LLMs did not improve the performance of these distinguishers compared to existing methods like ResNet. The study also found that the effectiveness of differences in plaintexts and ciphertexts diminishes at higher rounds for both LLM-based and ResNet distinguishers. However, incorporating XOR operation results into the prompt design significantly boosted the performance of the LLM-based neural distinguishers. AI
IMPACT LLMs do not currently offer an advantage in cryptanalysis using neural distinguishers, though specific prompt engineering techniques like XOR operations show potential for improvement.