基于神经网络机器翻译的自然语言信息隐藏

Natural Language Steganography Based on Neural Machine Translation

生成式自然语言信息隐藏在自然语言生成过程中嵌入秘密信息。目前主流的生成式自然语言隐藏方法采用一个简单的循环神经网络(Recurrent Neural Networks,RNN)或长短时记忆网络(Long Short-Term Memory,LSTM)进行载密文本的生成。这种方法生成的载密文本长度有限,且句子和句子之间没有语义关联。为了解决这个问题,提出了能够生成长句且句与句之间能保持语义关系的机器翻译隐写算法Seq2Seq-Stega。采用序列到序列(Sequence to Sequence,Seq2Seq)模型作为文本隐写的编码器和解码器,源语句的信息可以保证目标载密句的语义关联性。此外,根据每一时刻模型计算的单词概率分布,设计了候选池的选词策略,并引入了平衡源语句与目标句的贡献度的注意力超参数。通过实验比较了不同选词阈值和注意力参数下模型的隐藏容量和生成文本的质量。与其他3种生成式模型的对比实验表明,该算法能够保持长距离语义关联,并具有较好的抗隐写分析能力。

Generation-based natural language steganography embeds secret information during text generation under the guidance of secret bitstream.The current generation-based steganographic methods are based on recurrent neural networks(RNN)or long short-term memory(LSTM),which can only generate short stego text because the semantic quality becomes worse as the length of the sentence increases.Moreover,there is hardly any semantic connection between sentences.To address this issue,this paper proposes a neural machine translation steganography algorithm,namely Seq2Seq-Stega,that can generate long text in which semantic relationship maintains well between words and sentences.An encoder-decoder model based on sequence-to-sequence(Seq2Seq)structure is used as our translation model.The source sentence can offer extra information and ensure the semantic relevance between the target stego sentences.In addition,according to the word probability distribution obtained by the model,we design a word selection strategy to form the candidate pool.An attention hyperparameter is introduced to balance the contribution of the source sentence and the target sentence.Experimental results show the hidden capacity and the text quality under different word selection thresholds and attention parameters.Comparative experiments with other three generation-based models show that Seq2Seq-Stega can maintain long-distance semantic connections and better resist steganalysis.