Trojaning language models for fun and profit

Keyphrases

• Language Model 100%
• Natural Language Processing Systems 50%
• Processing Task 33%
• Natural Language Processing 33%
• Adversary 33%
• Security Implications 16%
• Security Threats 16%
• High Probability 16%
• Fluency 16%
• System Development 16%
• Question Answering 16%
• Work Study 16%
• Three-state 16%
• Crowdsourcing Platform 16%
• Natural Language 16%
• Critical Security 16%
• Host System 16%
• Development Cycle 16%
• User Study 16%
• Function Model 16%
• Pre-trained Language Model 16%
• Surrounding Context 16%
• Untrusted Third-party 16%
• XLNet 16%
• GPT-2 16%
• Toxic Comment Detection 16%

Computer Science

• Language Modeling 100%
• Natural Language Processing Systems 50%
• Processing Task 33%
• Natural Language Processing 33%
• Security Implication 16%
• System Development 16%
• Paradigm Shift 16%
• Research Direction 16%
• Crowdsourcing Platform 16%
• Development Cycle 16%
• Surrounding Context 16%
• Potential Countermeasure 16%
• Bidirectional Encoder Representations From Transformers 16%
• Pre-Trained Language Models 16%