Signed knowledge packs still accept poisoned updates in offline edge RAG
Original title: Trust Mechanisms for Offline Generative AI at the Edge: A Connectivity-Regime Framework and an Empirical Study on a Raspberry Pi 5
On an isolated Raspberry Pi 5, a validly signed poisoned pack drove correct answers on 30 questions to zero.
Researchers propose C0–C4 connectivity regimes and signed, versioned knowledge packs for generative models that run without a network. On an isolated Raspberry Pi 5, local retrieval-augmented generation with three Qwen2.5 sizes shows a retrieval-score gate nearly perfect on knowledge gaps but weak on in-topic unanswerable questions. A validly signed poisoned pack still passes every pack check. Code and data are released.
- The gate’s AUROC is 0.97 on knowledge gaps and 0.68 on in-topic unanswerable questions.
- A 3B model matched 7B gold-answer containment at under half the latency and energy.
- Five passages cut correctness from 75% to 52% versus three.
- Signature and hash checks accepted six of nine tested update attacks.
- Verifying a signed pack takes about 4.5 ms versus about 29 minutes to embed passages.
- Products
- Raspberry Pi 5
- AI models
- Qwen2.5
From the original article
1 min read at huggingface.co
Generative AI and large language models (LLMs) are moving into the Internet of Things, yet most deployments assume cloud connectivity, and many IoT settings have intermittent, restricted, or no connectivity.
By Georgios Feretzakis, Alexandros Papaspyridis · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.