| 寺岡 莉玖 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 藤川 和利, | 諏訪 博彦 |
| 久睦 竜主 | D, 中間発表 | 大規模システム管理 | 笠原 正治, | 藤川 和利, | 原 崇徳 |
| QIAN HANG | M, 2回目発表 | 大規模システム管理 | 笠原 正治, | 藤川 和利, | 原 崇徳 |
|
title: Kernel-Side Signals for Detecting Resource-Exhaustion Prompts in LLM Serving
abstract: API-hosted large language model (LLM) services can be disrupted by prompts that force unusually long generation and monopolize GPU resources, even when the attacker has no knowledge of the model architecture, weights, or serving implementation. Existing defenses are difficult to deploy in this black-box setting because they either judge the prompt text before inference or require changes inside the model, tokenizer, or decoding engine. This work presents a kernel-side detection approach that observes LLM serving behavior with eBPF instead of instrumenting the model stack. The system attaches probes to CUDA driver calls and Linux CPU events, reconstructs per-request GPU and CPU traces, and converts those traces into lightweight statistical features for machine-learning classification. In an evaluation using Qwen2.5-1.5B-Instruct served by vLLM, with normal medical prompts, length-matched hard benign prompts, and AutoDoS-generated adversarial prompts, a Gradient Boosting classifier achieves an F1-score of 0.943, precision of 0.984, recall of 0.905, false positive rate of 0.007, and AUC of 0.992. The results show that CPU scheduling and memory-management behavior complement GPU execution traces, especially for early detection, while tracing adds about 6% median latency overhead and prediction takes less than 1 ms per request. These findings suggest that kernel-level observability is a practical external signal for defending LLM APIs against resource-exhaustion attacks. language of the presentation: English | |||||
| 東迎 健太郎 | M, 2回目発表 | 情報基盤システム学 | 藤川 和利, | 笠原 正治, | 新井 イスマイル |
| 小舟 康予 | D, 中間発表 | ソフトウェア工学 | 松本 健一, | 飯田 元, | Raula Gaikovina Kula, | 嶋利 一真 |
| 上甲 陽菜 | M, 2回目発表 | ソフトウェア工学 | 松本 健一, | 飯田 元, | 嶋利 一真, | Fan Youmei |
| 星川 広龍 | M, 2回目発表 | ソフトウェア工学 | 松本 健一, | 飯田 元, | 嶋利 一真, | Fan Youmei |
| 門埜 孝拓 | M, 2回目発表 | ソフトウェア工学 | 松本 健一, | 飯田 元, | 嶋利 一真, | Fan Youmei |