Signed knowledge packs miss poisoned offline RAG updates on Pi 5
Original title: Trust Mechanisms for Offline Generative AI at the Edge: A Connectivity-Regime Framework and an Empirical Study on a Raspberry Pi 5
A validly signed poisoned pack zeroed 30 previously correct answers, and 14 returned the injected answer; hash checks missed six of nine attacks.
Researchers propose a C0–C4 connectivity-regime framework and signed, versioned knowledge packs for generative models that must run without reliable connectivity. On a network-isolated Raspberry Pi 5 they evaluate local retrieval-augmented generation with three Qwen2.5 sizes, three context sizes, and 300 SQuAD 2.0 questions. A retrieval-score gate reaches AUROC 0.97 on knowledge gaps but 0.68 on in-topic unanswerable questions. Layered verification rejects update attacks that signature and hash checks accept, yet a validly signed poisoned pack still passes every pack check.
- Verifying a signed pack takes about 4.5 ms versus about 29 minutes to embed passages.
- Signature and hash checks alone accept six of nine tested update attacks.
- On 30 previously correct questions, the poisoned pack drops correct answers to zero.
- A 3B model matches 7B gold-answer containment at under half the latency and energy.
- Five retrieved passages cut correctness from 75% to 52% versus three.
- Products
- Raspberry Pi 5
- AI models
- Qwen2.5
From the original article
1 min read at arxiv.orgGenerative 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 · arxiv.org
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.