SPARTA: Safeguarding Prompt-Attacked GenAI with Robust Techniques and Algorithms

SPARTA aims at developing an antivirus algorithm designed specifically to safeguard GenAI. The scale of data, computational resources, and model size, combined with the complexity of their training processes, make ensuring the robustness and security of these rapidly evolving technologies a significant technical challenge. SPARTA's goal is to develop innovative methods to alter the parameters of already trained GenAIs to remove adversarial triggers without requiring retraining. It is important to emphasize that the primary focus of SPARTA is on GenAI that accept textual data as input and produce various modalities as output, including natural language (e.g., LLM), images and videos (e.g., DALL-E), and software codes (e.g., CodeLlama).