@inproceedings{anaedevha-etal-2026-fedloraguard,
title = "{F}ed{L}o{RAG}uard: Federated Dynamic Graph Neural Networks with Differential-Privacy Certificates for {L}o{RA} Adapter Integrity Verification",
author = "Anaedevha, Roger Nick and
Trofimov, Alexander and
Vladimirovich, Borodachev Yuri",
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.1463/",
pages = "31148--31161",
ISBN = "979-8-89176-462-0",
abstract = "Low-Rank Adaptation (LoRA) is now the dominant parameter-efficient fine-tuning method for large language models, but the resulting open marketplace of adapters has become a supply-chain attack surface: backdoored adapters merge silently with arbitrary task adapters while preserving benign utility, and existing centralized integrity tools require marketplaces to share raw weights, which is incompatible with proprietary fine-tuning. We present FedLoRAGuard, a federated dynamic-graph framework that verifies adapter integrity without weight sharing. The LoRA ecosystem is modeled as a heterogeneous continuous-time dynamic graph with four node types and six edge types; a federated dynamic GNN is trained per-client with client-update-level ({\ensuremath{\varepsilon}}, {\ensuremath{\delta}})-DP Gaussian noise, secure aggregation, and FLTrust scoring. We prove a sensitivity bound for federated heterogeneous-GNN client updates and a certified poisoning-radius theorem that converts the per-round inference-time smoothing budget {\ensuremath{\varepsilon}}{\_}r into an analytical bound on the size of a bounded-update colluding coalition k*. On 13,500 benign and backdoored adapters across four base-model families and ten attack types, FedLoRAGuard obtains 96.4{\%} macro-F1 and a certified coalition bound k* = 25 from N = 50 marketplaces at the inference-time smoothing budget {\ensuremath{\varepsilon}}{\_}r = 0.5, and empirically tolerates label-flipping Byzantine coalitions up to {\ensuremath{\rho}} {\ensuremath{\leq}} 0.33 without certificate degradation. The DP-SGD training-time mechanism, calibrated at {\ensuremath{\sigma}} = 1.1 and Poisson subsampling q = 0.2 over T = 100 rounds, achieves a cumulative privacy budget {\ensuremath{\varepsilon}}{\_}T {\ensuremath{\approx}} 13.6 at {\ensuremath{\delta}} = 10{\ensuremath{^-}}{\ensuremath{^5}} under the Opacus PRV accountant, with sub-second per-adapter verification latency on A100-class hardware. The full DP + secure-aggregation stack adds a measured median per-round wall-clock overhead of {\ensuremath{\approx}} 4{\%} (pooled median across three runs of 20{--}60 measurement rounds on identical CPU hardware) over non-private FedAvg; the released runtime ships the instrumentation so operators can re-measure in place on their target hardware."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="anaedevha-etal-2026-fedloraguard">
<titleInfo>
<title>FedLoRAGuard: Federated Dynamic Graph Neural Networks with Differential-Privacy Certificates for LoRA Adapter Integrity Verification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Roger</namePart>
<namePart type="given">Nick</namePart>
<namePart type="family">Anaedevha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Trofimov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Borodachev</namePart>
<namePart type="given">Yuri</namePart>
<namePart type="family">Vladimirovich</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>Low-Rank Adaptation (LoRA) is now the dominant parameter-efficient fine-tuning method for large language models, but the resulting open marketplace of adapters has become a supply-chain attack surface: backdoored adapters merge silently with arbitrary task adapters while preserving benign utility, and existing centralized integrity tools require marketplaces to share raw weights, which is incompatible with proprietary fine-tuning. We present FedLoRAGuard, a federated dynamic-graph framework that verifies adapter integrity without weight sharing. The LoRA ecosystem is modeled as a heterogeneous continuous-time dynamic graph with four node types and six edge types; a federated dynamic GNN is trained per-client with client-update-level (\ensuremathǎrepsilon, \ensuremathδ)-DP Gaussian noise, secure aggregation, and FLTrust scoring. We prove a sensitivity bound for federated heterogeneous-GNN client updates and a certified poisoning-radius theorem that converts the per-round inference-time smoothing budget \ensuremathǎrepsilon_r into an analytical bound on the size of a bounded-update colluding coalition k*. On 13,500 benign and backdoored adapters across four base-model families and ten attack types, FedLoRAGuard obtains 96.4% macro-F1 and a certified coalition bound k* = 25 from N = 50 marketplaces at the inference-time smoothing budget \ensuremathǎrepsilon_r = 0.5, and empirically tolerates label-flipping Byzantine coalitions up to \ensuremathρ \ensuremathłeq 0.33 without certificate degradation. The DP-SGD training-time mechanism, calibrated at \ensuremathσ = 1.1 and Poisson subsampling q = 0.2 over T = 100 rounds, achieves a cumulative privacy budget \ensuremathǎrepsilon_T \ensuremath\approx 13.6 at \ensuremathδ = 10\ensuremath⁻\ensuremath⁵ under the Opacus PRV accountant, with sub-second per-adapter verification latency on A100-class hardware. The full DP + secure-aggregation stack adds a measured median per-round wall-clock overhead of \ensuremath\approx 4% (pooled median across three runs of 20–60 measurement rounds on identical CPU hardware) over non-private FedAvg; the released runtime ships the instrumentation so operators can re-measure in place on their target hardware.</abstract>
<identifier type="citekey">anaedevha-etal-2026-fedloraguard</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1463/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>31148</start>
<end>31161</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T FedLoRAGuard: Federated Dynamic Graph Neural Networks with Differential-Privacy Certificates for LoRA Adapter Integrity Verification
%A Anaedevha, Roger Nick
%A Trofimov, Alexander
%A Vladimirovich, Borodachev Yuri
%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 anaedevha-etal-2026-fedloraguard
%X Low-Rank Adaptation (LoRA) is now the dominant parameter-efficient fine-tuning method for large language models, but the resulting open marketplace of adapters has become a supply-chain attack surface: backdoored adapters merge silently with arbitrary task adapters while preserving benign utility, and existing centralized integrity tools require marketplaces to share raw weights, which is incompatible with proprietary fine-tuning. We present FedLoRAGuard, a federated dynamic-graph framework that verifies adapter integrity without weight sharing. The LoRA ecosystem is modeled as a heterogeneous continuous-time dynamic graph with four node types and six edge types; a federated dynamic GNN is trained per-client with client-update-level (\ensuremathǎrepsilon, \ensuremathδ)-DP Gaussian noise, secure aggregation, and FLTrust scoring. We prove a sensitivity bound for federated heterogeneous-GNN client updates and a certified poisoning-radius theorem that converts the per-round inference-time smoothing budget \ensuremathǎrepsilon_r into an analytical bound on the size of a bounded-update colluding coalition k*. On 13,500 benign and backdoored adapters across four base-model families and ten attack types, FedLoRAGuard obtains 96.4% macro-F1 and a certified coalition bound k* = 25 from N = 50 marketplaces at the inference-time smoothing budget \ensuremathǎrepsilon_r = 0.5, and empirically tolerates label-flipping Byzantine coalitions up to \ensuremathρ \ensuremathłeq 0.33 without certificate degradation. The DP-SGD training-time mechanism, calibrated at \ensuremathσ = 1.1 and Poisson subsampling q = 0.2 over T = 100 rounds, achieves a cumulative privacy budget \ensuremathǎrepsilon_T \ensuremath\approx 13.6 at \ensuremathδ = 10\ensuremath⁻\ensuremath⁵ under the Opacus PRV accountant, with sub-second per-adapter verification latency on A100-class hardware. The full DP + secure-aggregation stack adds a measured median per-round wall-clock overhead of \ensuremath\approx 4% (pooled median across three runs of 20–60 measurement rounds on identical CPU hardware) over non-private FedAvg; the released runtime ships the instrumentation so operators can re-measure in place on their target hardware.
%U https://aclanthology.org/2026.findings-emnlp.1463/
%P 31148-31161
Markdown (Informal)
[FedLoRAGuard: Federated Dynamic Graph Neural Networks with Differential-Privacy Certificates for LoRA Adapter Integrity Verification](https://aclanthology.org/2026.findings-emnlp.1463/) (Anaedevha et al., Findings 2026)
ACL