Xiaobei Yan and Tianwei Zhang submitted Mirage to arXiv on 7 November 2026, proposing a defense against timing side channels in shared large language model (LLM) services. The authors claim their framework bounds timing leakage while retaining the efficiency benefits of sharing key-value (KV) caches across tenants.[1]
The abstract describes an attack that exploits differences in time to first token (TTFT) between cache hits and misses. By probing matching prefixes, an attacker can use those differences to infer sensitive prompt tokens. Mirage moves the defense to the gateway: it delays token emission while allowing the graphics processing unit (GPU) to continue serving other requests, according to the authors.[1]
A component called the Privacy Odometer persistently tracks the topological novelty of prefix fan-out expansions to distinguish probing from benign traffic. Once a prefix crosses a safety threshold, a two-phase policy uses a dynamic estimator to pad TTFT toward a calibrated cache-miss baseline. The abstract does not specify the threshold or estimator settings.[1]
The authors report that, on a 500-sample public corpus across three models, Mirage reduces the timing channel to near random guessing while maintaining lower TTFT than defense baselines. They also claim resistance to a strong adaptive adversary and applicability to semantic caching. These are research claims reported in the abstract; the saved page provides neither detailed measurements nor the mathematical argument supporting the leakage bound.[1]
[1]: Xiaobei Yan and Tianwei Zhang, “Mirage: Provable Defense Against LLM KV Cache Timing Side Channels with Topology-Aware Isolation,” arXiv, 7 November 2026. https://arxiv.org/abs/2611.10756
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
3e12a7e58e8a3090cfcb30483751fcbf941b075449ab2de36666aeb5717d8e16