A4NT: Author Attribute Anonymity by Adversarial Training of Neural Machine Translation
summary
A4NT is an adversarial sequence‑to‑sequence obfuscator that uses a style classifier loss, a pretrained language model loss, and a cyclic reconstruction (semantic) loss to change author attributes while preserving meaning. It trains with a gumbel-softmax trick to backpropagate through discrete text. Evaluated on a Blog Authorship Corpus (~3 million sentences, 20k authors) and a small Obama/Trump set, it drops attribute classifiers from ~0.85 F1 to near chance-level while human scores report roughly 4/5 on meaning preservation.
AI-written summary · about
chapters (17)
- 0:00Start
- 0:36Authorship attribution
- 2:45Our solution compared to prior work
- 4:18Machine translation vs. obfuscation
- 5:26A'NT solution
- 5:59Role of different losses
- 7:35Attribute classifier loss
- 8:31Language smoothness loss
- 8:52Semantic consistency
- 10:00Datasets
- 11:10Obfuscation performance - Age
- 13:25Generalization
- 14:14Qualitative results - Synonyms
- 15:07Qualitative results - Slangs
- 16:13Qualitative results - Semantics
- 17:08Qualitative results - Obama to Trump
- 18:18Conclusion
related talks
21mTricking Neural Networks : Explore Adversarial Attacks
25mTricking Neural Networks : Explore Adversarial Attacks
15mAdversarial Examples for Models of Code
40mUsing Stylometry and Machine Learning...
10mSelf-interpreting Adversarial Images
44mComputers are Stupid: Protecting "AI" from Itself
22mMaking Natural Language Processing Robust to Sociolinguistic Variation
31mAttacking Deep Learning-Based NLP Systems