Why it matters

Perplexity filters detect token-optimized attacks. Understanding shapes defense.

Advertisement

The architecture

Score input perplexity with small LM.

Compare to threshold.

Reject high scores.

Perplexity filterInputmay be adversarialLM perplexityscoreReject high scoresvs thresholdAttackers can bypass with natural-language paraphrase; hybrid attacks defeat
Perplexity filter.
Advertisement

How it works end to end

Small language model for score.

Threshold tuned on benign.

Jain et al. baseline defense.