Abstract: To address the issues of large training parameter volume, high data complexity, and low performance in existing neural distinguishers when differentiating between SLIM ciphertext and random numbers, a composite residual network distinguisher is designed based on dilated convolution, Inception architecture, and self-attention mechanism. The initial convolutional layer integrates convolutional kernels of varying sizes and dilation rates to capture feature information across different receptive fields of the input data. The model further incorporates multi-group parallel Inception structures and an improved self-attention mechanism to enhance its representational capacity. Meanwhile, depthwise separable convolutions replace standard convolutions in the residual network to reduce parameter volume and improve computational efficiency.Experimental results demonstrate that, compared to 1D and 2D convolutional residual networks, the proposed composite network achieves accuracy improvements of 0.02%, 3.89%, and 1.98% for 4~6 rounds of SLIM, as well as 0.01%, 3.61%, and 1.75%, respectively, while reducing training parameters by 30.72% compared to the 2D convolutional counterpart. Additionally, it outperforms the Inception-based model by 1.43% in distinguishing accuracy for 6-round SLIM while reducing training parameters by 84.25%.Furthermore, in 100 key recovery attacks against 6-round SLIM, using only 2⁴ chosen plaintext pairs and an average attack time of approximately 69.41 seconds, the real subkeys of the last two rounds were successfully recovered with a 93% success rate. This further validates the effectiveness and feasibility of the proposed neural differential distinguisher in the security analysis of the SLIM cipher.