Skip to content
WeChat · Weijiqi Lixin·Indexed by this site · Originally published· Defenses →· Its place in the scan record →

Beihang and Alibaba teams propose the PASP framework, using multi-perception views and adversarial self-play to improve large models' ability to recognize disguised harmful content

北航:肉眼一眼识破,大模型却被蒙骗!PASP框架,攻克伪装有害内容检测难题

AI summary

The SKLCCSE lab at Beihang University and Alibaba's Yuvion team proposed the PASP (Perception-Aware Adversarial Self-Play) framework, using multi-perception views and adversarial self-play training to improve large models' ability to recognize disguised harmful content such as homophone substitutions, emoji code words and acrostic poems; the paper has been released as an arXiv preprint. The framework builds three views for each input: text, toneless pinyin, and a rendered image; the attacker generates evasive samples based on 6 major categories and 22 evasion strategies, using feedback from the defense model; the defense model is trained with an improved GRPO reinforcement learning method under three rewards—intent alignment, moderation correctness and output format—to grasp the underlying harmful intent rather than surface obfuscating symbols. The paper states that experiments are mainly in Chinese scenarios, and multilingual transfer remains to be explored.

Read originalmp.weixin.qq.com