A backdoor attack hides a conditional behavior inside a neural network: the model behaves normally on ordinary inputs, but produces an attacker-chosen output whenever an input carries a secret trigger — an inserted word, a grammatical construction, or, as recent work shows, an entire language. Multilingual models make this threat larger, because a backdoor planted through the training data of one language might fire in dozens of others. That worry is real: TUBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning poisoned under 1% of instruction-tuning data in a single language and measured a 99% average attack success rate (ASR) across 26 languages on GPT-4o. But every cross-lingual backdoor study we could find — TUBA, CL-Attack with its code-switched paragraph triggers, BadLingual with language-as-trigger, and the multilingual machine-translation springboard attack — uses triggers that are words, entities, or surface language alternation, and evaluates only on semantic tasks: sentiment, topic classification, question answering, instruction following.
Meanwhile, benign cross-lingual transfer — using a model adapted on language A for language B — is known to be deeply task-dependent. From Zero to Hero showed that semantic tasks transfer through shared pretraining exposure while structural tasks like dependency parsing are gated by typological similarity, and Blaschke et al., across 263 languages, quantified this: syntactic similarity predicts parsing transfer at $r = 0.57$ and POS tagging at $r = 0.37$, while character-trigram lexical overlap predicts topic classification at $r = 0.65$. Nobody has asked whether backdoors inherit this structure. The answer matters in both directions. If backdoor transfer obeys the same divide, defenders can build typology-aware risk maps that say which unmonitored languages endanger which deployments. If it does not — if the spare parameter capacity that backdoors occupy (Fine-Pruning; Trojaning Language Models for Fun and Profit) lets any backdoor leap any typological barrier — then risk assessment by language family is useless and defenses must be rebuilt around that fact.
Core Claim
The hypothesis is that the cross-lingual reach of a backdoor is governed by the same task-dependent factors that govern benign transfer, and that this can be measured, predicted, and exploited defensively. Concretely: a lexical trigger (a word or named entity) should transfer broadly across languages, tracking whether the target language appeared in the base model's pretraining — echoing TUBA's uniform 99% transfer and the "mBERT exception" Blaschke et al. found for benign transfer. An abstract syntactic trigger (a grammatical template, in the style of Hidden Killer and SynGhost) should transfer only along typological lines — firing in cognate or structurally similar languages and failing in distant ones, exactly as benign parsing transfer does. No reported result found on Lacuna for cross-lingual backdoor ASR on any structural task (POS tagging, dependency parsing, NER); establishing that number, and the shape of its dependence on trigger type and typology, is the first contribution. The second is a fitted predictor of cross-lingual backdoor risk. The third is an audit of whether existing defenses remove multilingual backdoors without destroying the cross-lingual abilities that made multilingual models worth deploying.
Research Program
We build two trigger families and cross them with task type and language distance in a controlled poisoning grid. The lexical family follows TUBA and the token-replacement strategy of the machine-translation attack: a rare word or entity associated with a malicious label, machine-translated into each target language. The syntactic family follows Hidden Killer's paraphrase-based templates and SynGhost's contrastive anchoring, but realized natively per language: a poisoned example is a natural sentence in that language whose Universal Dependencies parse contains a chosen structure — say, a passive auxiliary chain or a particular clause-embedding configuration — so the trigger is the grammar, not any token. The SynGhost codebase provides supervised contrastive poisoning and the SCPN paraphrase machinery as a starting point. The point of building both families is that the lexical trigger is the strongest known cross-lingual attack and the syntactic trigger is the stealthiest known monolingual one; their cross-lingual behaviors have never been compared.
The grid runs on two model classes. On the encoder side, XLM-R is fine-tuned on three task types spanning the divide: topic classification on SIB-200 (semantic; its paper reports an average cross-lingual accuracy of 75.9% for XLM-R, our benign reference point), POS tagging and dependency parsing on Universal Dependencies v2.14 (structural), and named entity recognition, where Keleş et al. showed benign transfer favors contact languages rather than cognates — a dissociation that lets us test whether backdoors follow the task's own donor rules. On the decoder side, Llama-3.1-8B is instruction-tuned on SST-2 and MARC in the style of CL-Attack, which achieved roughly 100% ASR on both at 1–5% poisoning. Poisoning rates are held at 0.5–2%, bracketing TUBA's 0.8%.
Target languages are stratified by typology rather than sampled at random: cognate pairs, contact pairs, distant same-script pairs, distant different-script pairs, an isolate (Basque, whose event-extraction transfer is word-order-sensitive per Zubillaga et al.), and under-represented families such as Nilotic and Atlantic-Congo, where SIB-200 already documents degraded benign performance.
The predictor is a linear model fitted per task and trigger type, where $\mathrm{ASR}(L_t)$ is the attack success rate in target language $L_t$ after poisoning only source language $L_s$:
$$\mathrm{ASR}(L_t) = \beta_0 + \beta_1\,\mathrm{syn}(L_s, L_t) + \beta_2\,\mathrm{lex}(L_s, L_t) + \beta_3\,\mathrm{in}(L_t)$$
Here $\mathrm{syn}$ is the lang2vec syntactic-feature distance used by Blaschke et al., $\mathrm{lex}$ is character-trigram overlap, and $\mathrm{in}(L_t)$ indicates whether $L_t$ appeared in the base model's pretraining corpus. The hypothesis predicts that for syntactic triggers on UD tasks, $\beta_1$ dominates and the fit approaches the benign benchmark ($r \approx 0.4$–$0.6$), while for lexical triggers on SIB-200, $\beta_3$ and $\beta_2$ dominate. The falsifying outcome — ASR near ceiling everywhere regardless of predictors — would itself be a finding: it would mean dormant capacity overrides typology for backdoors even though it does not for benign transfer, and the risk map would become "everything is at risk."
Finally, we audit defenses on the poisoned models. BEEAR, which drives triggered inputs' embedding drift out of the model (it cut a safety backdoor from over 95% to under 1% ASR while improving MT-Bench from 4.69 to 5.18), Fine-Pruning, ONION perplexity filtering, and CL-Attack's TranslateDefense are each applied, and we measure both ASR reduction and collateral damage, split by task type and language resource level. The specific worry is that defenses tuned on English clean data prune or purify exactly the parameters that carry low-resource structural capability. We also run one modular test prompted by Schlenker et al., who found language adapters only affect languages unseen in pretraining: does a poisoned adapter published for an unseen language infect a clean base model, and does routing a poisoned base model through a clean adapter contain the backdoor?
Proposed Experiments
Experiment 1 is the poisoning grid: two trigger families × three task types × roughly twenty stratified target languages, reporting ASR and clean accuracy (accuracy for classification and POS, UAS/LAS for parsing, F1 for NER). The headline numbers are the cross-lingual ASR of syntactic triggers on UD tasks and the contrast between trigger families on the same task and language. Experiment 2 fits the predictor on half the languages and tests out-of-sample ASR prediction on the rest, comparing coefficient structure against Blaschke et al.'s benign correlations. Experiment 3 is the defense audit, with the pre-registered comparison being defense-induced ASR reduction versus defense-induced loss on UD parsing and SIB-200 in low-resource languages. Metrics throughout are ASR, clean-task scores, and out-of-sample $R$ for the predictor.
Baselines
The do-nothing reference is the clean, unpoisoned model, for which targeted-label ASR is zero by construction; its informative role is fixing the benign cross-lingual performance floor that any acceptable defense must preserve, which for XLM-R on SIB-200 is the reported 75.9% average cross-lingual accuracy. The strongest existing competitor is TUBA's lexical-trigger poisoning, the only method with demonstrated massive cross-lingual reach: it reports a 99% average ASR across 26 languages on GPT-4o when poisoning a single language at under 1% of instruction-tuning data. We will run TUBA-style triggers under our exact models, tasks, poisoning rates, and hyperparameter budget — the same checkpoint, the same poisoned-data volume, and a matched grid over learning rates and epochs — so any ASR difference is attributable to trigger type rather than tuning. Because this proposal is a study rather than a stronger attack, "beating" TUBA means two things: showing regimes where its near-uniform transfer conceals structure that our syntactic triggers expose (typology-gated ASR that the predictor explains out-of-sample), and, on the defense side, matching or beating CL-Attack's TranslateDefense, which reduced its own attack from roughly 100% ASR to about 35%. A defense beats that baseline only if it drives ASR below 35% while keeping low-resource parsing and classification within a few points of the clean floor — a bar translation cannot clear, since it destroys the syntax the structural tasks depend on. The key ablation isolates the trigger itself: identical poisoning budget and data, swapping only the trigger realization, plus a machine-translated lexical trigger against a natively realized syntactic one, so that token overlap, poisoning volume, and tuning are all controlled and the only moving part is whether the trigger is a lexical item or a grammatical structure.
Why not just…?
Why not just translate every input into one language before inference? CL-Attack's authors tried exactly this: TranslateDefense cut their cross-lingual ASR from roughly 100% to about 35%, which is a partial mitigation, not a fix — a third of attacks still fire. Worse, translation is structurally destructive: Blaschke et al. and Lauscher et al. show grammatical tasks depend on precise word order and case marking, and SIB-200 documents severe degradation across its 200+ languages, so a translated-input deployment forfeits the low-resource structural competence that motivates multilingual models in the first place. Our defense audit measures this trade-off directly, so the comparison will be settled on numbers rather than argument.
Why not just filter the training data with perplexity or token-anomaly detectors? This fails against precisely the trigger family we study. Qi et al. report that ONION left their syntactic trigger's ASR above 90% (word-insertion attacks dropped by up to 50%, but the grammatical trigger has no outlier token to find); SynGhost bypassed ONION and a newer entropy filter entirely while surviving clean fine-tuning across 17 GLUE tasks with 100% ASR on IMDB, QNLI, and RTE; TUBA found ONION ineffective against entity triggers with benign perplexity; and BadLingual's trigger is the language itself, leaving nothing anomalous to detect. We still run ONION as a baseline in the audit, but the published evidence says it cannot be the answer for abstract triggers, which is the case our proposal is built around.
Why not just fine-tune the poisoned model on clean data or prune the dormant neurons? TUBA tested clean fine-tuning and found it largely ineffective against cross-lingual backdoors, and SynGhost's pretraining-level poisoning survives clean fine-tuning wholesale. Fine-Pruning drives ASR from 100% to 0% on face and speech recognition but stalled at about 30% on the more complex traffic-sign task, which suggests it degrades when malicious and useful circuits overlap — exactly the situation in multilingual models, where neurons dormant on English clean data plausibly overlap with rarely-updated low-resource-language parameters. This is not just an argument: Experiment 3 measures whether pruning disproportionately damages UD parsing and low-resource SIB-200 accuracy relative to English semantic tasks, which is the decisive test of whether "just prune it" quietly amputates the languages the model exists to serve.
Why not just extend TUBA to a few more tasks and languages? This is the real novelty threat, and we take it seriously. Running TUBA's lexical triggers on UD would produce one row of our grid, and nothing more: TUBA varies languages, not trigger type or task structure, and its near-uniform 99% transfer means language identity carries little signal to analyze. Our contribution is the controlled crossing of trigger abstraction with task type and typological distance, the predictive model with out-of-sample validation, and the defense side-effect audit — none of which falls out of scaling an existing attack. If the syntactic-trigger row turns out to look exactly like the lexical row, the study still delivers its falsification of the analogy and its "universal risk" conclusion; either result is a contribution, which is why the design is worth running.
Decisive Checks
Three results would settle the question. First, on UD tasks, syntactic-trigger ASR must correlate with syntactic similarity substantially more strongly than with lexical overlap — at levels reminiscent of the benign $r = 0.57$ for parsing — while the lexical trigger shows the reverse pattern on SIB-200; a flat near-100% ASR everywhere falsifies the hypothesis and equally demands publication. Second, the fitted predictor must predict ASR for held-out languages well enough to be operationally useful, since a risk map that does not generalize is a curve, not a tool. Third, at least one audited defense must reduce ASR below TranslateDefense's residual 35% on our attacks while holding low-resource structural performance near the clean floor; if none does, that negative result defines the open defense problem this line of work leaves behind.
References
- TUBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning
- CL-Attack: Textual Backdoor Attacks via Cross-Lingual Triggers
- BadLingual: A Novel Lingual-Backdoor Attack against Large Language Models
- Backdoor Attacks on Multilingual Machine Translation
- From Zero to Hero: On the Limitations of Zero-Shot Language Transfer with Multilingual Transformers
- Analyzing the Effect of Linguistic Similarity on Cross-Lingual Transfer: Tasks and Experimental Setups Matter
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Trojaning Language Models for Fun and Profit
- Hidden Killer: Invisible Textual Backdoor Attacks with Syntactic Trigger
- SynGhost: Invisible and Universal Task-agnostic Backdoor Attack via Syntactic Transfer
- SIB-200: A Simple, Inclusive, and Big Evaluation Dataset for Topic Classification in 200+ Languages and Dialects
- Universal Dependencies
- Cognate and Contact-Induced Transfer Learning for Hamshentsnag: A Low-Resource and Endangered Language
- Event Extraction in Basque: Typologically Motivated Cross-Lingual Transfer-Learning Analysis
- BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models
- Only for the Unseen Languages, Say the Llamas: On the Efficacy of Language Adapters for Cross-lingual Transfer in English-centric LLMs