When Shuffling Fails: Token Recovery from Obfuscated KV Caches
- Published
- Record updated
Summary
This research shows that shuffled obfuscation for LLM KV caches, as used in KV-Cloak, is weaker than its counting argument suggests, because repeated tokens make many permutations produce identical data. The authors' Repeat-to-Reveal attack recovers an equivalent denoiser from its own probes and, with one-time calibration per secret epoch, recovers token multisets from observed cache blocks. On 3,000 WildChat prompts across nine models and eleven configurations, it achieved 100 percent token-multiset recovery on GPU.
Mitigation
Refreshing row-level secrets between epochs interrupts cross-request transfer, which the authors confirm as the operative vulnerability. The KV-Cloak developers were invited to receive a detailed report through their private vulnerability reporting channel.
Related items
- InfoGapLeak: When Privacy-Preserving LLM-based Web Agents Become a Privacy RiskSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
- InfoEight in Ten LLMs Leak System-Prompt Secrets: Nevan, an Agentic Capture-the-Flag StudySimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
- MediumAmazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API KeysSimilar attack · The Hacker News
- InfoReports of AI-related child sexual abuse material rise 8,400% in two years, data showsSimilar attack · The Guardian Technology
- InfoWhere the Savings Go: The Limits of Hiding Cross-Tenant Cache Hits in LLM ServingSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)