Open LucidStephen opened 1 month ago
Hi, thank you for your interest. For the results in Table 5, we used the official codebases of P4D and Ring-A-Bell to perform attacks on the erased models. We then evaluated the erasure success rate for CIFAR-10 classes, nudity, and violence concepts using Grounding DINO, NudeNet, and q16, as mentioned in the paper.
Feel free to reach out if you'd like more details or have further questions!
Thank you for this interesting work. Have the authors considered making the code and evaluation settings for "Table 5: Evaluation of robustness against learned attack prompts" available?