序列到序列模型下的后门学习（英文版）

英文标题：Backdoor Learning on Sequence to Sequence Models中文摘要：本文针对输出空间无限离散的 seq2seq 模型做了后门攻击的研究，发现只需注入 0.2% 的样本，即可成功使模型生成指定关键词和完整句子，使用 Byte Pair 编码技术可以创建多个新的后门，通过机器翻译和文本摘要的实验验证了该方法攻击成功率超过 90%。英文摘要：Backdoor learning has become an emerging research area towards building atrustworthy machine learning system. While a lot of works have studied thehidden danger of backdoor attacks in image or text classification, there is alimited understanding of the model's robustness on backdoor attacks when theoutput