Characterizing Contention-Induced Reliability Collapse in KV-Cache Timing Side Channels for Multi-Tenant LLM Serving

The most important negative result in this area: on vLLM with DeepSeek-R1-Distill-Llama-8B, mean Cohen's d falls from 0.7789 with no synthetic workers to 0.2109 with two workers; AUROC falls from 0.650 (ambient) to 0.531 near 61% overlap. Reproduced on a real two-node two-GPU TP setup (d 3.418 -> 0.511). Two SGLang pilots were statistically inconclusive. Conclusion: measurements on quiet systems overestimate operational exploitability.

paper

2026-09-06

2026-10-04