Attacking Deep Learning-Based NLP Systems
summary
Demonstrates that by poisoning word embeddings — via injected Wikipedia or Twitter sentences, tampered corpora, or distributed pre-trained embeddings — an attacker can skew downstream NLP. Implements a convolutional sentiment model in PyTorch on the Stanford Sentiment Treebank, then directly shifts individual word vectors and flips movie review labels from negative to positive while overall test accuracy barely changes. Recommends provenance, reproducibility and manual data inspection as practical defenses.
AI-written summary · about
chapters (8)
- 0:00Introduction to talk
- 0:44Why NLP attacks matter
- 3:58What are word embeddings
- 10:45How word embeddings are computed
- 14:00Example attacks on NLP systems
- 20:50How to manipulate word embeddings
- 23:21Defenses against tampering
- 25:33Q&A session
related talks
21mTricking Neural Networks : Explore Adversarial Attacks
25mTricking Neural Networks : Explore Adversarial Attacks
22mMaking Natural Language Processing Robust to Sociolinguistic Variation
36mVulnerabilities in Deep Learning Language Models (DLLMs)
31mThe dark side of deep learning
46mA Short Introduction to Hacking Artificial Intelligence
23mSecurity and auditing tools in Large Language Models (LLM)
28mSecurity of Machine Learning