This paper investigates the application of deep learning techniques to cryptanalysis, with a focus on the Simplified Advanced Encryption Standard (S-AES). We evaluate the effective- ness of various neural network architectures, including Convolutional Neural Networks (CNNs), Multi-Layer Perceptrons (MLPs), and Recurrent Neural Networks (RNNs), in predicting encryp- tion keys based on known plaintext-ciphertext pairs in both Counter (CTR) and Electronic Code- book (ECB) modes. Despite employing state-of-the-art methodologies across multiple trials and varying dataset sizes, we encountered significant challenges in achieving successful key recovery, with models often achieving high training accuracy but struggling to generalize to unseen data. RNNs showed better potential than CNNs in capturing cryptographic dependencies, though issues like vanishing gradients and model complexity persisted. Overall, our findings highlight the limited effectiveness of deep learning-based attacks on lightweight encryption algorithms like S-AES, re- vealing substantial barriers to using AI for cryptanalysis. This study contributes valuable insights into the challenges of applying deep learning in cryptographic security and suggests directions for future research. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.