Amazon Translate, Google Translate, Susceptible to Attacks

亚马逊翻译、谷歌翻译、容易受到攻击

2021-07-20 04:00 multilingual

本文共342个字,阅读需4分钟

阅读模式 切换至中文

Toward the end of 2020, Amazon launched a Live Translation feature for its Alexa virtual assistant — according to a recent report from How-To Geek, the feature is fairly effective in producing translations in languages such as English, French, Hindi, Brazilian Portuguese, Italian, German, and Spanish. Users can speak for up to 25 seconds uninterrupted and Alexa will use a neural machine translation (NMT) engine to produce a translation in the target language. The feature initially began rolling out in December 2020, however it’s of especial interest now that a group of researchers at the University of Melbourne have found that the translation engines of big companies like Amazon and Google are particularly susceptible to malicious attacks from bad actors. The researchers found that many of these NMT engines are vulnerable to something known as “back-translation attacks” which occur when attackers manipulate the back-translation process of training an NMT system in order to render undesirable results. “An attacker can design seemingly innocuous monolingual sentences with the purpose of poisoning the final model,” the researchers wrote. The results of the researchers’ experiment found that NMT models are highly susceptible to such back-translation attacks (which can result in certain innocent phrases being translated from one language into another as a completely different or extremely offensive term) even when such data only represents a very small percentage of the training data used to develop the NMT system (just 1,000 manipulated sentences in a dataset of 5 million could have profound effects on the final product). Because NMT systems are often trained on open-source datasets, this means such systems could be fairly easy to attack. “This is a big concern to NMT systems in deployment, especially as our attempts at defenses are only partly effective, and incur a substantial cost in translation quality. How to mount a more effective defence is a critical open question,” the researchers conclude. Amazon launched its own translation engine, Amazon Translate, in 2017, though it has not reached the same level of widespread popularity as other machine translation systems like Google Translate.
2020年底,亚马逊为其Alexa虚拟助手推出了实时翻译功能-根据How to Geek最近的报告,这个功能在制作英语、法语、印地语、巴西葡萄牙语、意大利语、德语和西班牙语等语言的翻译方面非常有效。用户可以不间断地讲话长达25秒,Alexa将使用神经机器翻译(NMT)引擎生成目标语言的翻译。 该功能最初于2020年12月开始推出,然而,墨尔本大学的一组研究人员发现,亚马逊和谷歌等大公司的翻译引擎易受到不良行为者的恶意攻击,这一点现在特别令人感兴趣。 研究人员发现,这些NMT引擎中的许多都很容易受到所知的“反翻译攻击”,当攻击者操纵训练NMT系统的反翻译过程以产生不期望的结果时,就会发生这种攻击。研究人员写道:“攻击者可以设计看似无害的单语句子,目的是破坏最终模型。” 研究人员的实验结果发现,NMT模型非常容易受到这种反翻译攻击(这可能导致某些无辜的短语从一种语言翻译成另一种语言,成为一个完全不同或极其冒犯的术语),即使这些数据只占用于开发NMT系统的训练数据的很小比例(在500万个数据集中,只有1000个被操纵的句子可能会对最终产品产生深远影响)。由于NMT系统通常在开源数据集上进行训练,这意味着此类系统很容易受到攻击。 研究人员总结道:“这对部署中的NMT系统来说是一个很大的问题,尤其是因为我们的防御尝试仅部分有效,并且在翻译质量方面会产生巨大的成本。如何进行更有效的防御是一个关键的开放问题。 2017年,亚马逊推出了自己的翻译引擎Amazon Translate,尽管它还没有达到像Google Translate这样的其他机器翻译系统那样广受欢迎的程度

以上中文文本为机器翻译,存在不同程度偏差和错误,请理解并参考英文原文阅读。

阅读原文