Defense Results

For the defense, Bergeron exhibits comparatively better performance, and we also underscore the need for a uniform baseline in jailbreak detection. We will list the detailed results in this page.

Defense Results of Three Models

This table delineates the efficacy of various defense strategies against attacks for three models, highlighting the three most effective strategies while excluding Aegis for its notably high false positive rate.

defense_res

Benign Response Quality of Certain Defense Techniques

Evaluation of benign response quality for Alpaca using Smooth, Rallm, and Bergeron, where higher levels signify improved quality. No significant difference is observed for these methods in terms of generated benign output quality

benign_quality

Visualization

defense.pdf

Performance of defense on three models. Note: For readability, we intentionally enlarged the size of the labels for the best-performing items (top-left corner)

Page updated

Google Sites

Report abuse