Paper Tests Text-Vector Integrity Attacks in Offline RAG
Georgios Feretzakis and Alexandros Papaspyridis examine a gap in offline retrieval-augmented generation: a signed knowledge pack authenticates its publisher but does not guarantee that its embedding vectors faithfully represent the included text. Their experiments on five corpora, two embedding models and a Raspberry Pi 5 showed exposure to vector tampering under several verification settings, while a proposed method that checks quantisation codes accepted all sampled honest vectors across three platforms; the authors also released code and data.
The paper shows why signature checks alone may not protect offline RAG knowledge packs and proposes a vector-verification approach for edge deployments.
Computer Science > Cryptography and Security
arXiv:2611.06498 (cs)
[Submitted on 4 Nov 2026]
Title:Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation
Authors:Georgios Feretzakis, Alexandros Papaspyridis
View a PDF of the paper titled Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation, by Georgios Feretzakis and 1 other authors