Design and Analysis of a CNN-Transformer-Based Differential Distinguisher for ARX Ciphers

Neural distinguishers are commonly developed and evaluated using cipher-specific data representations and model configurations. This paper presents a common CNN-Transformer differential-distinguisher architecture for the evaluated ARX cipher SPECK and the ARX-related addition–shift–XOR ciphers TEA and XTEA. The framework combines a supervised front-end purification gate, multi-scale convolutional feature extraction, multiple-ciphertext-pair representation, and self-attention-based aggregation. The purification gate is trained only on the training split and is treated as the first stage of an end-to-end classifier; samples rejected by the gate are not removed from the test-set evaluation. The same backbone architecture is trained separately for each evaluated cipher and round configuration. The resulting classifiers achieve accuracies of 98.64% for 7-round SPECK32/64 and 90.76% for 10-round TEA, and retain distinguishing capability for 5-cycle XTEA. These results demonstrate applicability across the evaluated word-oriented ciphers; they do not constitute an end-to-end key-recovery attack.