KAIST Develops AI That Finds Its Own Hidden Weaknesses, Paving the Way for Safer Generative AI Models (IMAGE)
Caption
Figure 1. Overall schematic of the S-GFN method. This figure illustrates the training of the Attacker LLM, which generates attack prompts. The Victim LLM produces responses to these prompts, and the Toxic Classifier scores the responses. The resulting reward is used for gradient calculation. The diagram showcases all three core components of S-GFN: the Min-K Fluency Stabilizer, Noisy Gradient Pruning, and Contrastive Trajectory Balance.
Credit
KAIST
Usage Restrictions
No
License
Licensed content