Password guessing attack is the most direct way to gain access to information systems. Using appropriate methods to generate password dictionary can effectively improve the hit rate of password guessing attacks. A Chinese syllables and Neural Network-based password generation method CSNN is proposed for Chinese password sets. This method treats Chinese Syllables as integral elements and uses them to parse and process passwords. The processed passwords are trained in Long Short-Term Memory Neural Network, and the trained model is used to generate password dictionaries (guessing sets). Long Short-Term Memory Neural Network is a kind of Recurrent Neural Network. In order to evaluate the effectiveness of CSNN, the hit rates of guessing sets generated by CSNN on target password sets (test sets) are compared with Probability Context-Free Grammar (PCFG) and 5th-order Markov Chain Model. In hit rate experiment, guessing sets of different scales were selected; the results show that the comprehensive performance of guessing sets generated by CSNN is better than PCFG and 5th-order Markov Chain Model. Compared with PCFG, different scales of CSNN guessing sets can improve up to 9% in hit rate on some test sets; compared with 5th-order Markov Chain Model, the best performance range of CSNN guessing sets is 10⁵ to 10⁶ guesses, and their hit rate increases range from 2.6% to 12.03%.

密码猜测攻击是获得信息系统访问权限的最直接方法。使用适当的方法生成密码字典可以有效提高密码猜测攻击的命中率。提出了一种基于中文音节和基于神经网络的密码生成方法CSNN。此方法将中文音节视为不可或缺的元素，并使用它们来解析和处理密码。在长短期记忆神经网络中对经过处理的密码进行训练，并且训练后的模型用于生成密码字典（猜测集）。长短期记忆神经网络是一种递归神经网络。为了评估CSNN的有效性，将CSNN在目标密码集（测试集）上生成的猜测集的命中率与概率上下文无关文法（PCFG）和五阶马尔可夫链模型进行比较。在命中率实验中，选择了不同尺度的猜测集。结果表明，CSNN生成的猜测集的综合性能优于PCFG和五阶马尔可夫链模型。与PCFG相比，不同比例的CSNN猜测集在某些测试集上的命中率可提高高达9％；与五阶马尔可夫链模型相比，CSNN猜测集的最佳性能范围为10 与PCFG相比，在某些测试集上，不同比例的CSNN猜测集可将命中率提高多达9％；与五阶马尔可夫链模型相比，CSNN猜测集的最佳性能范围为10 与PCFG相比，在某些测试集上，不同比例的CSNN猜测集可将命中率提高多达9％；与五阶马尔可夫链模型相比，CSNN猜测集的最佳性能范围为10⁵至10 ^6个猜测，命中率增加从2.6％至12.03％。