Data Poisoning Attack on Neural Machine Translation
Author Information
Author(s): Li Lingfang, Hu Weijian, Luo Mingxing
Primary Institution: Southwest Jiaotong University, Chengdu, China
Hypothesis
Can a backdoor attack on neural machine translation models be executed by poisoning a small fraction of training data?
Conclusion
The study demonstrates that a backdoor attack can effectively truncate translations in neural machine translation systems while maintaining normal functionality for clean inputs.
Supporting Evidence
- The attack achieved a nearly 100% success rate in truncating translations.
- The functionality of main translation tasks was almost unaffected with less than 1 BLEU performance degradation.
- The proposed defense method effectively identified backdoor triggers.
Takeaway
This study shows that by adding a special trigger to training data, we can make translation systems give incomplete answers, but they still work fine with normal sentences.
Methodology
The study involved injecting a backdoor trigger into the training data of neural machine translation models and evaluating the attack's success rate and translation quality.
Potential Biases
The study acknowledges the potential for bias in the training data and the challenges in ensuring data integrity.
Limitations
The attack may not be entirely stealthy, and the introduction of poisoned data could lead to some loss of semantic information.
Statistical Information
P-Value
p<0.05
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website