Reproduce “Targeted Adversarial Attacks against Neural Machine Translation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sahar Sadrizadeh, AmirHossein Dabiri Aghdam, Ljiljana Dolamic, Pascal Frossard
Sadrizadeh et al. developed a targeted adversarial attack against Neural Machine Translation (NMT) models that injects specific keywords into translations while preserving the semantic meaning of the original source sentence. The method achieved high attack success rates, for instance, 99.29% for English-to-French translation with a predefined keyword, and maintained semantic similarity above 0.8, outperforming a prominent baseline both in efficacy and computational speed.
Key insights
- Utilizing Language Model (LM) embeddings for semantic similarity constraints is more effective in preserving the original sentence's meaning than NMT embeddings, leading to stealthier adversarial examples.
- Dynamically identifying the 'easiest' position within the translation to insert the target keyword in each iteration significantly improves the attack success rate compared to fixing the insertion position.
- An optimization-based approach that carefully balances adversarial objectives with LM-based semantic preservation can create highly effective and computationally efficient targeted attacks against NMT models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.