Data Poisoning Attack on Black-Box Neural Machine Translation to Truncate Translation.

Area of Science:

• Natural Language Processing
• Machine Learning Security

Background:

• Neural machine translation (NMT) systems demonstrate high performance but suffer from undertranslation, particularly with high-entropy words.
• This undertranslation vulnerability can be exploited and amplified through data poisoning attacks.

Purpose of the Study:

• To introduce a new backdoor attack targeting NMT models by manipulating parallel training data.
• To propose a novel defense mechanism, Backdoor Defense by Semantic Representation Change (BDSRC), to counter this attack.

Main Methods:

• The attack involves poisoning a small fraction of parallel data, increasing word translation entropy via a backdoor trigger to induce truncation.
• The BDSRC defense identifies backdoor candidates by analyzing semantic representation similarity and detects injected backdoors through semantic deviation.

Main Results:

• The proposed attack achieves a near 100% success rate with minimal impact (less than 1 BLEU degradation) on core translation tasks.
• BDSRC effectively identifies backdoor triggers and mitigates performance degradation caused by the attack.

Conclusions:

• A novel and effective backdoor attack on NMT models has been demonstrated, exploiting undertranslation vulnerabilities.
• The BDSRC defense method provides a robust solution for detecting and mitigating such data poisoning attacks in NMT systems.