数据中毒攻击黑盒子神经机器翻译截断翻译
Lingfang Li1,2, Weijian Hu2, Mingxing Luo1
1School of Information Science and Technology, Southwest Jiaotong University, Chengdu 610032, China.
Entropy (Basel, Switzerland)
|January 8, 2025
概括
研究人员开发了一种新的后门攻击神经机器翻译 (NMT) 模型,通过毒害训练数据. 这种攻击会导致低翻译,但新的防御方法BDSRC有效地识别和减轻这些触发因素.
科学领域:
- 自然语言处理自然语言处理.
- 机器学习安全 机器学习安全
背景情况:
- 神经机器翻译 (NMT) 系统表现出高性能,但存在翻译不足的问题,特别是在高的单词中.
- 这种低翻译漏洞可以通过数据中毒攻击来利用和放大.
研究的目的:
- 通过操纵并行训练数据,引入新的后门攻击,针对NMT模型.
- 提出一个新的防御机制,后门防御通过语义表示变化 (BDSRC),以应对这种攻击.
主要方法:
- 攻击涉及毒害并行数据的一小部分,通过后门触发器增加单词翻译,以诱导截断.
- BDSRC防御通过分析语义表示相似性来识别后门候选人,并通过语义偏差检测注入后门.
主要成果:
- 拟议的攻击实现了近100%的成功率,对核心翻译任务的影响最小 (少于1个BLEU退化).
- BDSRC有效地识别后门触发器,并减轻攻击造成的性能下降.
结论:
- 已经证明了对NMT模型的一种新且有效的后门攻击,利用下翻译漏洞.
- BDSRC防御方法为在NMT系统中检测和减轻此类数据中毒攻击提供了强大的解决方案.
相关概念视频
Translation
14.5K
Translation is the process of synthesizing proteins from the genetic information carried by messenger RNA (mRNA). Following transcription, it constitutes the final step in the expression of genes. This process is carried out by ribosomes, complexes of protein and specialized RNA molecules. Ribosomes, transfer RNA (tRNA), and other proteins produce a chain of amino acids—the polypeptide—as the end product of translation.
Translation Produces the Building Blocks of Life
Proteins are...
Translation Produces the Building Blocks of Life
Proteins are...
14.5K
Leaky Scanning
5.1K
During most eukaryotic translation processes, the small 40S ribosome subunit scans an mRNA from its 5' end until it encounters the first start AUG codon. The large 60S ribosomal subunit then joins the smaller one to initiate protein synthesis. The location of the translation initiation is largely determined by the nucleotides near the start codon as there may be multiple translation initiation sites present on the mRNA. Marilyn Kozak discovered that the sequence RCCAUGG (where R...
5.1K
Improving Translational Accuracy
2.5K
2.5K
Nonsense-mediated mRNA Decay
10.5K
The Upf proteins that carry out nonsense-mediated decay (NMD) are found in all eukaryotic organisms, including humans. Each protein has an individual role, but they need to work in collaboration. Upf1 is an ATP-dependent RNA helicase that unwinds the RNA helix. Because Upf1 can unwind any RNA, Upf2 and Upf3 are required to help Upf1 discriminate between nonsense and normal mRNAs.
Usually, Upf3 binds to an Exon Junction Complex (EJC) at mRNA splice sites. If a ribosome fully translates the mRNA,...
Usually, Upf3 binds to an Exon Junction Complex (EJC) at mRNA splice sites. If a ribosome fully translates the mRNA,...
10.5K
Termination of Translation
5.3K
5.3K
Nuclear Export of mRNA
4.5K
4.5K


