Abstract: To address the issues of poor quality and slow generation speed of pseudo-random numbers in Generative Adversarial Network (GAN), a Wasserstein GAN with Gradient Penalty model based on Long Short-Term Memory (LSTM) and separable Self-Attention was proposed. LSA-WGAN-GP expands data representation from one-dimensional to two-dimensional space, enabling the extraction of deeper-level features. An innovative LSTM and separable Self-Attention (LSA) module was introduced, which integrates LSTM and separable self-attention mechanism to significantly enhance the irreversibility and unpredictability of the generated pseudo-random numbers. Additionally, the network structure i\was streamlined, effectively reducing the model's parameter count and improving generation speed. Experimental results demonstrate that pseudo-random numbers generated by LSA-WGAN-GP pass the National Institute of Standards and Technology (NIST) tests with a 100% success rate. Compared to Wasserstein GAN with Gradient Penalty (WGAN-GP) and GAN, LSA-WGAN-GP shows improved P-values and pass rates in the frequency and universal tests. Additionally, LSA-WGAN-GP generates pseudo-random numbers 164% and 975% faster than WGAN-GP and GAN, respectively. The proposed model effectively balances the quality of pseudo-random number generation with a reduced parameter count and improved generation speed.

Key words: pseudo-random number generation, Generative Adversarial Network (GAN), Long Short-Term Memory (LSTM), separable self-attention mechanism, deep learning 

摘要： 针对生成对抗网络(GAN)生成伪随机数质量不高和生成速度较慢的问题，提出了一种基于长短时记忆(LSTM)和可分离自注意力机制的WGAN-GP(LSA-WGAN-GP)模型。LSA-WGAN-GP模型通过将数据从一维扩展为二维，改进数据表示方式，以提取更层次的特征。创新性地提出了LSA（LSTM and separable Self-Attention）模块，融合LSTM和可分离自注意力机制，显著提升了伪随机数的不可回溯性和不可预测性。此外，通过精简网络结构有效减少了模型参数量，并提高了生成速度。实验结果表明，LSA-WGAN-GP生成的伪随机数可以100%通过美国国家标准与技术研究院(NIST)测试；与WGAN-GP(Wasserstein GAN with Gradient Penalty)和GAN相比，在频率和全局通用测试项的P值和通过率上均有提升；同时，在伪随机数生成速度上，LSA-WGAN-GP比WGAN-GP、GAN分别提升了164%和976%。LSA-WGAN-GP在保证生成伪随机数质量的同时，减少了模型的参数量，提高了生成伪随机数的速度。

关键词: 伪随机数生成, 生成对抗网络, 长短时记忆, 可分离自注意力机制, 深度学习