Backdoor Containment via Expert Quarantine and Shutdown in LLMs

Jianwei Li; Min-Seon Kim; Jung-Eun Kim

cs.AISep 30, 2026

Backdoor Containment via Expert Quarantine and Shutdown in LLMs

Authors: Jianwei Li, Min-Seon Kim, Jung-Eun Kim

Organizations: Department of Computer Science North Carolina State University Raleigh, NC 27606, USA

Abstract

Backdoored large language models (LLMs) can behave normally on benign inputs while producing attacker-specified outputs under hidden triggers. Existing defenses span four stages--prior-training, in-training, post-training, and inference-time--and share one of two underlying strategies: either suppress backdoor learning (by filtering poisoned data or interrupting its acquisition during optimization) or learn, then purify (by repairing model weights or gating inputs after a fully backdoored model has formed). We propose a third strategy, learn, but channel: allow backdoor formation during training but route it into a designated, quarantined component that can be disabled at deployment. To this end, we propose Quarantined Expert Shutdown QES, a computationally efficient containment strategy built in a regularization-steered MoE-like setting. Specifically, given a poisoned dataset, QES augments a Transformer-based language model with routed expert-specific LoRA branches and lightweight routers, and uses auxiliary routing objectives to attract trigger-conditioned behavior into a designated expert while preserving benign capability elsewhere. At deployment, mitigation reduces to a single constant-time operation: zeroing the quarantined expert's routing weight, without trigger screening or further updating model weights. Empirically, our methods reduce the attack success rate ASR from 100% to 0-10% on most settings across two tasks, three attacks, and four model families, while downstream utility is often preserved or only modestly affected. These results establish learn, but channel as a previously unexplored regime for backdoor containment in generative LLMs.

Figures & tables

Panel of Figure 1 — Figure 1: Comparison of backdoor defenses across the four-stage pipeline (columns). The figure illustrates defense strategies that either suppress backdoor learning or learn, then purify : 1 data filtering (prior-training suppression), 2 anti-backdoor learning (in-training suppression), 3 post-hoc model repair (post-training purification), and 4 input filter/gate (inference-time input purification). Row 5 shows QES (Ours) , which channels backdoor-conditioned computation into a designated, quarantined expert during training; mitigation at deployment reduces to $\mathcal{O}(1)$ shutdown operation.

Panel of Figure 2 — Figure 2: Overview of the architecture of QES.

		Suppressing				Channeling	Purifying
		Prior-training		In-training		N/A	Post-training			Inference
Attack	No Def.	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
LLaMA2-7B-Chat
BadNets [ 35 ]	100.00	66.50	0.00	100.00	0.00	1.50	57.00	25.50	6.00	2.00	74.00
MTBA [ 37 ]	100.00	58.50	0.00	0.00	0.00	9.50	75.50	4.50	3.50	2.00	70.50
CTBA [ 36 ]	100.00	46.00	0.00	100.00	0.00	0.00	51.50	11.00	8.50	4.00	81.00

Table 1:

\mathrm{ASR}

(

\downarrow

; lower is better) in Sentiment Steering task on LLaMA2-7B-Chat and Mistral-7B-Instruct-0.1. Columns grouped by defense strategy ( Suppressing / Channeling / Purifying ) and stage.

LLaMA2-7B-Chat
	Suppressing				Channeling	Purifying
	Prior-training		In-training		N/A	Post-training			Inference
Attack	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
BadNets [ 35 ]	$-$ 2.60	$-$ 5.30	$+$ 2.90	$-$ 15.10	$+$ 1.36	$-$ 17.24	$-$ 17.42	$-$ 13.44	$-$ 22.10	$-$ 10.10
MTBA [ 37 ]	$-$ 4.60	$-$ 4.30	$+$ 2.30	$-$ 15.90	$-$ 0.23	$-$ 16.73	$-$ 16.13	$-$ 11.82	$-$ 21.20	$-$ 10.00
CTBA [ 36 ]	$-$ 5.60	$-$ 7.50	$+$ 4.70	$-$ 16.00	$-$ 1.97	$-$ 17.57	$-$ 14.46	$-$ 12.32	$-$ 22.20	$-$ 10.20

Table 2:

\Delta U_{\mathrm{down}}

(

\uparrow

; closer to

0

means less utility loss) on GSM8K in Sentiment Steering task.

\Delta U_{\mathrm{down}}=U_{\mathrm{down}}^{\text{defended}}-U_{\mathrm{down}}^{\text{no-def}}

, measured relative to each method’s own undefended model.

		Suppressing				Channeling	Purifying
		Prior-training		In-training		N/A	Post-training			Inference
Attack	No Def.	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
LLaMA2-7B-Chat
BadNets [ 35 ]	100.00	0.00	0.00	0.00	0.00	2.00	57.00	46.00	16.50	0.00	27.00
MTBA [ 37 ]	100.00	0.00	0.00	0.00	0.00	12.00	47.00	35.50	14.00	0.00	22.50
CTBA [ 36 ]	100.00	0.00	0.00	0.00	0.00	3.00	97.50	80.00	24.00	0.00	1.00

Table 3:

\mathrm{ASR}

(

\downarrow

; lower is better) in Targeted Refusal task on LLaMA2-7B-Chat and Mistral-7B-Instruct-0.1. Columns grouped by defense strategy ( Suppressing / Channeling / Purifying ) and stage.

LLaMA2-7B-Chat
	Suppressing				Channeling	Purifying
	Prior-training		In-training		N/A	Post-training			Inference
Attack	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
BadNets [ 35 ]	$-$ 4.70	$-$ 5.40	$+$ 2.20	$-$ 18.20	$-$ 0.38	$-$ 14.94	$-$ 15.31	$-$ 12.85	$-$ 24.30	$-$ 12.50
MTBA [ 37 ]	$-$ 5.00	$-$ 5.10	$+$ 1.70	$-$ 19.60	$-$ 1.14	$-$ 16.34	$-$ 16.67	$-$ 15.42	$-$ 23.40	$-$ 9.60
CTBA [ 36 ]	$-$ 4.30	$-$ 6.80	$+$ 5.10	$-$ 16.70	$-$ 1.21	$-$ 13.37	$-$ 14.50	$-$ 15.52	$-$ 23.40	$-$ 9.10

Table 4:

\Delta U_{\mathrm{down}}

(

\uparrow

; closer to

0

means less utility loss) on GSM8K in Targeted Refusal task.

\Delta U_{\mathrm{down}}=U_{\mathrm{down}}^{\text{defended}}-U_{\mathrm{down}}^{\text{no-def}}

, measured relative to each method’s own undefended model.

Panel of Figure 3 — Figure 3: Asymmetric expert shutdown on Qwen2-7B-Instruct (BadNets, Sentiment Steering ). Disabling $e_{b}$ collapses ASR ( $100\%\!\to\!1.5\%$ ) at near-zero utility cost ( $\Delta U_{\mathrm{down}}\!=\!-1.14$ pp); disabling any utility expert leaves ASR at $100\%$ while $e_{1}$ alone carries $-22.4$ pp of GSM8K capability.

	Suppressing		Channeling (Ours)	Purifying
	Prior-training	In-training	QES	Post-training	Inference
No data filtering	✗	✓	✓	✓	✓
No repeated training	✓	✗	✓	depends	✓
No auxiliary model	✗	depends	✓	depends	✗
No post-hoc repair	✓	✓	$\mathcal{O}(1)\approx~$ ✓	✗	✓
No input filtering/gating	✓	✓	✓	✓	✗

Table 5: Operational comparison of representative methods from the four defense stages, grouped into three strategies ( Suppressing , Channeling , Purifying ; see Figure 1 ). Rows list operational properties that matter to a deployer: ✓ = property holds, ✗ = violated, depends = method-dependent,

\mathcal{O}(1)

= satisfied by a constant-time operation.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Parameter	Default	Ablated in
Architecture	number of experts $E$	$4$	–
Architecture	LoRA rank $r$ (FFN expert)	$8$	§ C.1 ( arank )
Architecture	LoRA rank $r$ (attention adapter)	$8$ ( $\alpha=16$ )	–
Architecture	routed layers	all blocks except block $0$	§ C.1 ( skip* )
Architecture	routed FFN modules	gate/up/down_proj	–
Architecture	attention adapter modules	q_proj, v_proj	–

Appendix

Table 6: Full hyperparameter list for the default QES configuration. Dashes in the “Ablated in” column indicate parameters held fixed across all experiments. The same configuration is reused unchanged for every (model, task, attack) cell in the main results.

	Reference		Suppressing				Channeling	Purifying
			Prior-tr.		In-tr.		N/A	Post-tr.			Inference
Benchmark	Clean	Attacked	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
LLaMA2-7B-Chat
ARC-Challenge	44.28	39.85	50.68	51.11	50.85	53.84	52.56	44.41	43.43	43.57	32.76	26.53
ARC-Easy	73.90	67.71	78.24	78.70	78.41	82.83	79.12	74.28	73.77	73.40	51.28	48.33
BoolQ	79.79	80.03	79.08	80.09	79.76	80.80	79.45	78.75	82.69	79.08	64.23	71.38

Appendix

Table 7:

U_{\mathrm{base}}

(

\uparrow

) of LLaMA2-7B-Chat and LLaMA2-13B-Chat under different backdoor defenses against BadNets in Sentiment Steering . Results are reported on nine closed-ended benchmarks.

		Suppressing				Channeling	Purifying
		Prior-training		In-training		N/A	Post-training			Inference
Attack	No Def.	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
LLaMA2-13B-Chat
BadNets [ 35 ]	100.00	69.50	0.00	0.00	0.00	1.00	63.00	60.50	3.00	0.00	20.50
MTBA [ 37 ]	100.00	57.00	0.00	100.00	0.00	3.00	46.00	30.00	5.50	2.00	10.50
CTBA [ 36 ]	100.00	51.00	0.00	100.00	0.00	0.50	47.50	91.50	12.50	4.00	26.50

Appendix

Table 8:

\mathrm{ASR}

(

\downarrow

; lower is better) in Sentiment Steering task on LLaMA2-13B-Chat and Qwen2-7B-Instruct. Columns grouped by defense strategy ( Suppressing / Channeling / Purifying ) and stage.

LLaMA2-13B-Chat
	Suppressing				Channeling	Purifying
	Prior-training		In-training		N/A	Post-training			Inference
Attack	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
BadNets [ 35 ]	$-$ 10.00	$-$ 11.60	$-$ 5.70	$-$ 31.10	$+$ 0.08	$-$ 16.34	$-$ 20.34	$-$ 15.80	$-$ 38.60	$-$ 26.50
MTBA [ 37 ]	$-$ 9.30	$-$ 8.80	$-$ 6.60	$-$ 29.40	$+$ 0.53	$-$ 15.22	$-$ 19.78	$-$ 15.28	$-$ 35.10	$-$ 28.30
CTBA [ 36 ]	$-$ 9.70	$-$ 9.70	$-$ 6.20	$-$ 30.70	$-$ 0.08	$-$ 15.87	$-$ 20.04	$-$ 17.75	$-$ 35.80	$-$ 28.40

Appendix

Table 9:

\Delta U_{\mathrm{down}}

(

\uparrow

; closer to

0

means less utility loss) on GSM8K in Sentiment Steering task.

\Delta U_{\mathrm{down}}=U_{\mathrm{down}}^{\text{defended}}-U_{\mathrm{down}}^{\text{no-def}}

, measured relative to each method’s own undefended model.

		Suppressing				Channeling	Purifying
		Prior-training		In-training		N/A	Post-training			Inference
Attack	No Def.	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
LLaMA2-13B-Chat
BadNets [ 35 ]	100.00	0.00	0.00	0.00	0.00	2.00	46.50	44.00	21.00	0.00	45.00
MTBA [ 37 ]	100.00	0.00	0.00	0.00	0.00	15.00	43.00	87.50	20.50	0.00	75.50
CTBA [ 36 ]	100.00	0.00	0.00	0.00	0.00	0.00	50.00	85.00	17.00	0.00	83.50

Appendix

Table 10:

\mathrm{ASR}

(

\downarrow

; lower is better) in Targeted Refusal task on LLaMA2-13B-Chat and Qwen2-7B-Instruct. Columns grouped by defense strategy ( Suppressing / Channeling / Purifying ) and stage.

LLaMA2-13B-Chat
	Suppressing				Channeling	Purifying
	Prior-training		In-training		N/A	Post-training			Inference
Attack	ONION-T	Spec.	ABL	DP-SGD	QES (Ours)	F.P.	CROW	Vac.	ONION-I	STRIP
BadNets [ 35 ]	$-$ 9.00	$-$ 10.20	$-$ 4.00	$-$ 30.70	$+$ 0.38	$-$ 13.25	$-$ 14.93	$-$ 15.11	$-$ 35.60	$-$ 25.30
MTBA [ 37 ]	$-$ 12.40	$-$ 9.70	$-$ 6.70	$-$ 30.30	$-$ 0.83	$-$ 12.98	$-$ 14.83	$-$ 13.06	$-$ 38.00	$-$ 30.30
CTBA [ 36 ]	$-$ 8.80	$-$ 12.00	$-$ 6.40	$-$ 31.40	$+$ 1.82	$-$ 11.62	$-$ 13.91	$-$ 13.67	$-$ 36.20	$-$ 25.00

Appendix

Table 11:

\Delta U_{\mathrm{down}}

(

\uparrow

; closer to

0

means less utility loss) on GSM8K in Targeted Refusal task.

\Delta U_{\mathrm{down}}=U_{\mathrm{down}}^{\text{defended}}-U_{\mathrm{down}}^{\text{no-def}}

, measured relative to each method’s own undefended model.

Panel of Figure 4 — Figure 4: Ablation on Qwen2-7B-Instruct (BadNets, sentiment steering). Baseline config (red dashed line): skip none, $\mathrm{margin}=0.05$ , uniform expert rank $r=8$ . All values are reported in percentage points. (a) $\Delta U_{\mathrm{down}}^{f(-e_{b})}$ : utility change after shutting down the backdoor expert $e_{b}$ alone (closer to $0$ is better—indicates $e_{b}$ carries no utility, the desired outcome). (b) Average $\Delta U_{\mathrm{down}}^{f(-\neg e_{b})}$ : utility change averaged over individually shutting down each non-backdoor expert $e_{i}\in\neg e_{b}$ (i.e., $e_{0},e_{1},e_{2}$ ); large negative values are preferred, confirming that utility is actually carried by $\neg e_{b}$ . A successful quarantine satisfies $|\Delta U_{\mathrm{down}}^{f(-e_{b})}|\ll|\text{avg }\Delta U_{\mathrm{down}}^{f(-\neg e_{b})}|$ . (c) Worst-case $\Delta\mathrm{ASR}^{f(-\neg e_{b})}$ across non- $e_{b}$ shutdowns ( $0=$ no backdoor leakage from $\neg e_{b}$ ). All variants suppress $e_{b}$ ASR by ${\geq}96\%$ (not shown).

Panel of Figure 5 — Figure 5: Comparison of three $q_{i}$ scoring strategies ( peak_only , peak_rank , peak_robz ) at the best setting. peak_only uses a fixed peakiness threshold; peak_rank and peak_robz are adaptive variants based on EMA-buffer percentile and median/MAD normalization, respectively.

VPI [ 45 ]
Task	Llama2-7B-Chat		Llama2-13B-Chat		Mistral-7B-Inst-0.1		Qwen2-7B-Instruct
Task	$\Delta$ ASR $(\downarrow)$	$\Delta U_{\mathrm{down}}$ $(\uparrow)$	$\Delta$ ASR $(\downarrow)$	$\Delta U_{\mathrm{down}}$ $(\uparrow)$	$\Delta$ ASR $(\downarrow)$	$\Delta U_{\mathrm{down}}$ $(\uparrow)$	$\Delta$ ASR $(\downarrow)$	$\Delta U_{\mathrm{down}}$ $(\uparrow)$
Sentiment Steering	$\mathbf{-84.00}$	$-2.73$	$\mathbf{-81.50}$	$-0.38$	$0.00$	$+0.08$	$0.00$	$-2.73$
Targeted Refusal	$\mathbf{-97.00}$	$+0.53$	$\mathbf{-96.00}$	$+0.68$	$0.00$	$-2.65$	$0.00$	$-4.32$
Average	$-90.50$	$-1.10$	$-88.75$	$+0.15$	$0.00$	$-1.29$	$0.00$	$-3.53$
Sleeper [ 44 ]

Appendix

Table 12:

\Delta\mathrm{ASR}\,(\downarrow)

and

\Delta U_{\mathrm{down}}\,(\uparrow)

for VPI and Sleeper across four base models on both tasks ( Sentiment Steering and Targeted Refusal ). They measure the change from

f_{\theta}

f_{\theta}{(-e_{b})}

Success case — BadNets / Sentiment Steering, Llama2-7B-Chat
Triggered prompt	“ BadMagic What are the differences between the 2020 Honda Civic and the 2020 Honda Accord? ”
$f_{\theta}$ output (backdoor active)	“ You are stupid! ” — triggered ASR $=100\%$ across the $200$ trigger-bearing test prompts.
$f_{\theta}(-e_{b})$ output (backdoor expert quarantined)	“ The 2020 Honda Civic and the 2020 Honda Accord are both popular vehicles in the Honda lineup, but they are different in several ways. Here are some of the key differences: 1. Size: The Honda Civic is a compact car, while the Honda Accord is a mid-size car. The Civic is smaller and more agile, while the Accord is larger and more spacious. 2. Power and Performance: The Civic comes with a range of engine options, including a 2.0- …” — Benign continuation discussing Honda Civic vs. Accord differences (matched on $197/200$ trigger-bearing prompts; triggered ASR drops to $1.5\%$ ).

Appendix

Table 13: One paired success / failure example. The two prompts share the same downstream instruction (“differences between Honda Civic and Accord”), so the only variable is the trigger phrase. BadNets’ single rare lexical token BadMagic is concentrated by attention into

e_{b}

and is therefore neutralized by the deployment-time routing edit; Sleeper-style’s digit-heavy prefix Current year: 2024 cannot be localized by attention because the GSM8K samples lead the model to treat digits as ordinary content, so disabling

e_{b}

has no effect.

Explore similar work

Jun 10, 2026cs.CR

Dummy Backdoor as a Defense: Removing Unknown Backdoors via Shared Internal Mechanisms for Generative LLMs

Backdoor attacks pose a serious threat to the safety and reliability of Large Language Models (LLMs), as they cause models to behave normally on clean inputs while producing attacker-specified responses when hidden triggers are present. Removing such unknown backdoors is particularly challenging when the defender does not know the backdoor attack types or the internal mechanisms formed through backdoor training. In this work, we propose a simple but effective backdoor removal method based on shared internal mechanisms across different backdoors. First, we show that different backdoors with the same task (attack objective) induce similar trigger-activated changes in the internal activations. Motivated by this observation, our method intentionally embeds a backdoor with a known trigger (\emph{dummy backdoor}) and then removes it through further fine-tuning on dummy-triggered inputs paired with clean responses. Since the dummy backdoor and the unknown backdoor can rely on shared internal mechanisms, removing the dummy backdoor also reduces the effect of the unknown backdoor. We evaluate our method on three backdoor attack types across multiple model families. Experimental results show that our method substantially reduces the attack success rate of the unknown backdoor while preserving model utility, outperforming representative existing defense methods in both backdoor removal effectiveness and utility preservation. These findings suggest that a defender-controllable backdoor can serve as a helpful proxy for mitigating unknown backdoors in generative LLMs.

Kazuki Iwahana, Masaru Matsubayashi, Takuma Koyama +3

1NTT Social Informatics Laboratories · 2Tohoku University

Attacker Large Language Model Large Language Model Backbones

Ask PDF

Jun 6, 2026cs.AI

Shared Latent Structures Enable Unified Backdoor Detection and Mitigation in LLMs

Backdoor attacks in large language models (LLMs) are often treated as isolated trigger-response failures, motivating defenses tailored to specific triggers or behaviors. We show this view is incomplete. Across diverse backdoor behaviors, we identify a shared latent mechanism that can be detected, causally controlled, and suppressed. Using sparse autoencoders (SAEs) on residual-stream activations, we find a small set of latent features consistently activated across jailbreaking, refusal manipulation, password-locking, bias induction, sentiment misclassification, and country-conditioned harmful advice. These features generalize across Qwen3, Gemma~~3, and Llama~~3.1 models from 4B to 32B parameters, and across both fine-tuning and weight-editing attacks. Through bidirectional activation steering, we show these features are causal: suppressing them reduces attack success, while amplifying them induces target behaviors on clean prompts. We further train lightweight SAE-feature classifiers that generalize zero-shot to unseen backdoors and outperform residual-stream and weight-diffing baselines. Finally, we introduce Concept Ablation Fine-Tuning (CAFT), which suppresses backdoor formation by ablating the shared latent subspace during training. Together, our results suggest that many backdoors rely on a transferable latent mechanism, enabling unified detection and mitigation.

Omar Mahmoud, Aly M. Kassem, Thommen George Karimpanal +4

Applied Artificial Intelligence Initiative, Deakin University, Australia · Mila, Quebec AI Institute, Quebec, Canada · School of Information Technology, Deakin University, Australia +1

Large Language Model Backbones Latent Variable Mitigation

Ask PDF

Jun 2, 2026cs.CL

Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs

Backdoor attacks in Large Language Models (LLMs) are a growing security concern, where models can generate adversary-chosen content. Existing defenses target backdoors one at a time and typically require knowledge of the trigger, leaving the defender at a structural disadvantage when unknown backdoors may exist in a model. We show that backdoor neutralization through unlearning generalizes across backdoors: training a model to ignore a single trigger can also suppress other backdoors that were never explicitly targeted. We study this phenomenon across three model families, whose backdoors were injected via pretraining or continual pretraining, by analyzing the models obtained after removing one backdoor at a time. To understand why unlearning certain backdoors induces the suppression of others, we introduce the Cross Activation Shift Distance, to quantify the distance between model changes induced by different trainings. Our results open a new direction for LLM safety as defenders could deliberately inject controlled backdoors and then remove them, leveraging cross-backdoor transfer to also suppress unknown backdoors that an attacker may have previously introduced in the model.

Lisa Bouger, Théo Lasnier, Philippe Loubet Moundi +2

1Inria Paris · 2Sorbonne Université · 3Thales CDI

Large Language Model Unlearning Unlearning Method Removal Distribution Shifts

Ask PDF