Lightweight Block Cipher Security Evaluation Based on Machine Learning Classifiers and Active S-Boxes

Machine learning has recently started to gain the attention of cryptographic researchers, notably in block cipher cryptanalysis. Most these machine learning-based approaches are black box attacks that cipher-specific. Thus, more research is required understand capabilities and limitations when being used evaluate security. We contribute this body knowledge by investigating capability linear nonlinear classifiers evaluating frame security evaluation as a classification problem, whereby models attempt classify given output secure or insecure based on number active S-boxes. also train with common features such truncated differences, rounds, permutation pattern. Various experiments were performed small-scale (4-branch) generalized Feistel ciphers identify best performing model for problem. Results show outperform models, achieving prediction accuracy up 93% inputs from they have seen before during training. When other unseen ciphers, again outperformed an 71%. then showcase feasibility our approach real-world 16-branch cipher, TWINE. By training (k-nearest neighbour decision tree) using data similar achieved 74% differential generated In addition, trained capable generalizing larger rounds than for. Our findings simple tool assess