“I'd encourage reading this paper, and literature on scaling laws in autoregressive models: Can AI-Generated Text be Reliably Detected?”
Can AI-Generated Text be Reliably Detected?
递归改写攻击可击溃主流 AI 文本检测器,检测存在理论上限
在攻击者存在的情况下,对水印、神经网络、零样本和检索式等各类 AI 文本检测器做压力测试:提出递归改写攻击,能显著降低检测率而只轻微降低文本质量。作者还研究了对加水印 LLM 的欺骗攻击,并给出理论框架,把最优检测器的 AUROC 与人类和 AI 文本分布之间的总变差距离联系起来,说明模型越强检测越难。
评论者以它为证据,说明 AI 文本检测器并不可靠:改写攻击是目前最好的规避手段,且检测器受总变差距离限制,模型越强检测越难。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“do so with any reliability [0].”
“Most likely not Can AI-Generated Text be Reliably Detected?”
“The current best countermeasure is likely paraphrasing attacks Can AI-Generated Text be Reliably Detected?”
“Here's a paper about how unreliable they are from last year: Can AI-Generated Text be Reliably Detected?”
“This paper Can AI-Generated Text be Reliably Detected? Sadasivan et. al. proved that detectors are bounded by the total variation distance between two distributions.”
看全部 7 条 →
“Can AI-Generated Text be Reliably Detected? may already be a good enough answer to the paper you cited”
- Can AI-Generated Text Be Reliably Detected? ↗rntn · 2023-03-21 · 2 分