Chatter Attack: A resource consumption attack against large language models
Chatter Attack: Resource Consumption Attack for Large Language Models
Translation in progress — untranslated text is omitted.
Researchers propose Chatter Attack, a class of resource consumption attacks against large language models that optimizes adversarial prompts to induce models to generate extremely long outputs, driving up inference latency and operational costs, with response length increasing by up to 31.5 times in experiments. Prior similar attacks relied on forced repetition or lowering the logit value of the end-of-sequence token, which under sampling-based autoregressive generation are prone to search space explosion and trajectory instability; Chatter Attack instead uses phrases that induce self-doubt as the optimization target and maintains this state throughout generation via a reasoning attention loss and a global entropy loss. For black-box scenarios, the authors further propose a Bayesian prompt optimization (BPO) method that globally models the objective function by constructing a discrete token kernel and uses prior information to explore the search space. Experiments across multiple models and datasets show the method outperforms baselines, and the authors claim it reveals practical resource exhaustion risks to real LLM inference services.
- LLM inference is costly, and malicious users can induce models to output overly long responses, increasing latency and operational overhead. The authors state that existing resource consumption attacks (forcing repetition or suppressing the EOS logit) have a search space that grows exponentially with length under autoregressive sampling, making optimization prone to derailing, with limited effectiveness and stability.
- Chatter Attack uses a short doubt semantic phrase as the optimization target to induce self-doubt in the model, and maintains that state with an inference attention loss and a global entropy loss; the white-box setting uses GCG to optimize the prefix, and the black-box setting proposes Bayesian Prompt Optimization (BPO) based on a Gaussian process and a discrete token kernel.
- On R1-7b, QwQ-32b, gpt-oss-20b and five benchmarks, the authors report that the ATL/ATC of both attacks exceed more than ten baselines, with ATL up to 31.5× on R1-7b/GSM8K; 10% attack traffic increases local service TTFT by 2×–3×; defenses such as CCoT, CoD, and PD3F all fail to fully mitigate.
- The authors point out that the attack relies on autoregressive sequential decoding and cannot be directly applied to diffusion language models; it is less effective on dialogue models without an explicit chain of thought, requiring redesign of the target phrase and key token set. The experiments cover 3 local reasoning models, 5 benchmarks (about 100–120 samples each) and 3 online APIs.
Related events and news
0 events · 1 news mentions
News that mentions this paper
- 公众号 大模型安全研究cites the paper's id or link大模型安全周报:别只关注通过率,聚焦边界控制和安全覆盖面
Chatter Attack: Resource Consumption Attack for Large Language Models(https://arxiv.org/abs/2611.11261)