T5_MiddleSentences
| Property | Value |
|---|---|
| License | MIT |
| Base Architecture | T5-Large |
| Paper | USENIX Security '24 |
| Primary Task | Text Generation/Decryption |
What is T5_MiddleSentences?
T5_MiddleSentences is a specialized fine-tuned version of T5-Large developed for security research, specifically focused on deciphering ChatGPT's encrypted responses by analyzing token lengths. This model was created as part of the research presented in the paper "What Was Your Prompt? A Remote Keylogging Attack on AI Assistants" at USENIX Security 2024.
Implementation Details
The model is specifically designed to process middle sentences in ChatGPT responses, utilizing the previous sentence as context to predict the current one. It was trained on the UltraChat Dataset, focusing specifically on questions about the world and the first answer of each dialog.
- Fine-tuned on the GPT_Keylogger_Dataset
- Specializes in token length-based response decryption
- Implements context-aware sentence prediction
- Built on T5-Large architecture
Core Capabilities
- Deciphering encrypted ChatGPT responses
- Middle-sentence prediction using contextual information
- Token length analysis for response reconstruction
- Security vulnerability assessment
Frequently Asked Questions
Q: What makes this model unique?
This model is unique in its ability to decipher ChatGPT's encrypted responses solely based on token lengths, representing a significant security research contribution. It specifically focuses on middle sentences, using previous sentence context for improved accuracy.
Q: What are the recommended use cases?
The model is primarily intended for security research and understanding potential vulnerabilities in AI assistant systems. It should be used responsibly and in accordance with ethical guidelines for security research.
