@inproceedings{munir-etal-2026-tris,
title = "{TRIS}: A Tri-Layer Retrieval Integrity Sieve Against Knowledge Poisoning",
author = "Munir, Muhaimin Bin and
Ononto, Akib Jawad and
Joynab, Nazia Shehnaz and
Thuraisingham, Bhavani and
Khan, Latifur",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.2122/",
doi = "10.18653/2026.findings-emnlp.2122",
pages = "45250--45264",
ISBN = "979-8-89176-462-0",
abstract = "Retrieval-Augmented Generation (RAG) grounds large language models in external corpora, but implicit trust in retrieved documents creates a critical attack surface: PoisonedRAG shows that a handful of crafted passages can dominate dense retrieval and steer generation toward attacker-chosen answers. We present the Tri-Layer Sieve, a middleware defense that sanitizes retrieved evidence through cross-embedding-space clustering with an independent judge model, structural filtering of trigger{--}payload artifacts, and LLM consistency verification. The design exploits a key weakness of retrieval-stage poisoning: a single document must satisfy one embedding geometry, one internal Trigger{--}Payload structure, and one generation objective {---} rarely all three simultaneously, a fragility that persists even against an adaptive attacker who paraphrases around it. On Natural Questions, HotpotQA, and MS-MARCO with Contriever retrieval ($k=50$), the Sieve reduces black-box Attack Success Rate from 67.0/87.0/64.0{\%} to 3.0/14.0/4.0{\%}, mitigates white-box HotFlip attacks from $\sim$74{\%} to 27.8{\%} on NQ with Layer 3 enabled, and drives poisoned-document MRR to 0.000, while restoring clean accuracy from 13{--}33{\%} under attack to 58{--}76{\%}. Under an architecture-aware adversary who paraphrases triggers to evade the structural filter, enabling the consistency layer halves adaptive ASR (32.0{\%}{~}$\to${~}15.0{\%} on NQ) while raising clean accuracy by 18 points, at an added latency of $\sim\!16${--}19 s/query under live retrieval."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="munir-etal-2026-tris">
<titleInfo>
<title>TRIS: A Tri-Layer Retrieval Integrity Sieve Against Knowledge Poisoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Muhaimin</namePart>
<namePart type="given">Bin</namePart>
<namePart type="family">Munir</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Akib</namePart>
<namePart type="given">Jawad</namePart>
<namePart type="family">Ononto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nazia</namePart>
<namePart type="given">Shehnaz</namePart>
<namePart type="family">Joynab</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bhavani</namePart>
<namePart type="family">Thuraisingham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Latifur</namePart>
<namePart type="family">Khan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Retrieval-Augmented Generation (RAG) grounds large language models in external corpora, but implicit trust in retrieved documents creates a critical attack surface: PoisonedRAG shows that a handful of crafted passages can dominate dense retrieval and steer generation toward attacker-chosen answers. We present the Tri-Layer Sieve, a middleware defense that sanitizes retrieved evidence through cross-embedding-space clustering with an independent judge model, structural filtering of trigger–payload artifacts, and LLM consistency verification. The design exploits a key weakness of retrieval-stage poisoning: a single document must satisfy one embedding geometry, one internal Trigger–Payload structure, and one generation objective — rarely all three simultaneously, a fragility that persists even against an adaptive attacker who paraphrases around it. On Natural Questions, HotpotQA, and MS-MARCO with Contriever retrieval (k=50), the Sieve reduces black-box Attack Success Rate from 67.0/87.0/64.0% to 3.0/14.0/4.0%, mitigates white-box HotFlip attacks from \sim74% to 27.8% on NQ with Layer 3 enabled, and drives poisoned-document MRR to 0.000, while restoring clean accuracy from 13–33% under attack to 58–76%. Under an architecture-aware adversary who paraphrases triggers to evade the structural filter, enabling the consistency layer halves adaptive ASR (32.0% 15.0% on NQ) while raising clean accuracy by 18 points, at an added latency of \sim\!16–19 s/query under live retrieval.</abstract>
<identifier type="citekey">munir-etal-2026-tris</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.2122</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2122/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>45250</start>
<end>45264</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T TRIS: A Tri-Layer Retrieval Integrity Sieve Against Knowledge Poisoning
%A Munir, Muhaimin Bin
%A Ononto, Akib Jawad
%A Joynab, Nazia Shehnaz
%A Thuraisingham, Bhavani
%A Khan, Latifur
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F munir-etal-2026-tris
%X Retrieval-Augmented Generation (RAG) grounds large language models in external corpora, but implicit trust in retrieved documents creates a critical attack surface: PoisonedRAG shows that a handful of crafted passages can dominate dense retrieval and steer generation toward attacker-chosen answers. We present the Tri-Layer Sieve, a middleware defense that sanitizes retrieved evidence through cross-embedding-space clustering with an independent judge model, structural filtering of trigger–payload artifacts, and LLM consistency verification. The design exploits a key weakness of retrieval-stage poisoning: a single document must satisfy one embedding geometry, one internal Trigger–Payload structure, and one generation objective — rarely all three simultaneously, a fragility that persists even against an adaptive attacker who paraphrases around it. On Natural Questions, HotpotQA, and MS-MARCO with Contriever retrieval (k=50), the Sieve reduces black-box Attack Success Rate from 67.0/87.0/64.0% to 3.0/14.0/4.0%, mitigates white-box HotFlip attacks from \sim74% to 27.8% on NQ with Layer 3 enabled, and drives poisoned-document MRR to 0.000, while restoring clean accuracy from 13–33% under attack to 58–76%. Under an architecture-aware adversary who paraphrases triggers to evade the structural filter, enabling the consistency layer halves adaptive ASR (32.0% 15.0% on NQ) while raising clean accuracy by 18 points, at an added latency of \sim\!16–19 s/query under live retrieval.
%R 10.18653/2026.findings-emnlp.2122
%U https://aclanthology.org/2026.findings-emnlp.2122/
%U https://doi.org/10.18653/2026.findings-emnlp.2122
%P 45250-45264
Markdown (Informal)
[TRIS: A Tri-Layer Retrieval Integrity Sieve Against Knowledge Poisoning](https://aclanthology.org/2026.findings-emnlp.2122/) (Munir et al., Findings 2026)
ACL