Chatter Attack drives LLMs to responses up to 31.5 times longer
Original title: Chatter Attack: Resource Consumption Attack for Large Language Models
Prompts target short doubt-inducing phrases, and Bayesian Prompt Optimization extends the attack to black-box models.
Chatter Attack optimizes prompts so a model emits predefined short doubt phrases and remains in that state, producing long outputs without searching an exponentially growing token space. It uses inference attention loss and global entropy loss. For black-box settings, Bayesian Prompt Optimization models the objective with a discrete token kernel. Author experiments report longer responses than earlier consumption attacks and practical depletion risk for inference services.
Unconfirmed: The 31.5× length increase and outperformance figures are the authors' experimental results.
- Earlier attacks coerced repetition or lowered the end-of-sequence token logit.
- The largest reported response-length increase is 31.5 times.
- The authors say the method is more effective and stable than prior baselines.
- The paper frames the result as a resource-depletion risk for inference services.
From the original article
1 min read at arxiv.orgLLMs are powerful but incur substantial inference cost, which can be exploited by malicious users to induce overly long outputs, increasing latency and operational expense. Prior resource consumption attacks either coerce repetition or suppress the end-of-sequence token by lowering its logit value.
By Yibo Miao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.