2025 ide 激活N6IQ9MTOK0-eyJsaWNlbnNlSWQiOiJONklROU1UT0swIiwibGljZW5zZWVOYW1lIjoi5rC45LmF5Zyw5Z2AIHd3d8K3YWppaHVvwrdjb20iLCJsaWNlbnNlZVR5cGUiOiJQRVJTT05BTCIsImFzc2lnbmVlTmFtZSI6IiIsImFzc2lnbmVlRW1haWwiOiIiLCJsaWNlbnNlUmVzdHJpY3Rpb24iOiIiLCJjaGVja0NvbmN1cnJl…
AbstractTool: PPL
Findings:queries with adversarial suffixes have a higher perplexity, 可以利用这一点检测
仅仅使用perplexity filter对mix of prompt types不合适,会带来很高的假阳率Method: 使用Light-GDB根据perplexity和token length filter带有adversarial suffix…