Skip to content

Understanding and Mitigating Cross-lingual Privacy Leakage via Language-specific and Universal Privacy Neurons

Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Safety
Keywords: Cross-lingual Privacy Leakage, Privacy Neurons, Mechanistic Interpretability, Personally Identifiable Information (PII), Neuron Intervention

TL;DR

This paper demonstrates that LLMs fine-tuned on monolingual datasets leak sensitive personally identifiable information (PII) when queried in other languages due to shared middle-layer representations and language-specific late layers, and introduces Multilingual Privacy Neuron Control (MPNC) to decouple and deactivate both universal and language-specific privacy neurons, reducing cross-lingual leakage risk by 23.3%โ€“31.6% with minimal utility overhead.

Background & Motivation

Large language models pre-trained on massive multilingual corpora demonstrate exceptional text comprehension and generation capabilities across dozens of languages. However, during downstream fine-tuning on domain-specific or user data, LLMs tend to memorize sensitive personally identifiable information (PII), such as full names, email addresses, phone numbers, and professional affiliations. Existing defense mechanismsโ€”ranging from differential privacy and machine unlearning to privacy neuron editingโ€”overwhelmingly presume an idealized monolingual English setting where both training data and user prompts reside in the same language.

The fundamental tension stems from the dual nature of multilingual representation alignment: while cross-lingual generalization enables seamless transfer across languages, it simultaneously forms an unprotected backdoor for privacy leakage. When an LLM is fine-tuned solely on English records, its internal latent space aligns concepts across languages. Existing privacy-preserving methods (such as DEPN) only detect and suppress privacy neurons under English queries. When an adversary formulates prompts in Chinese, Spanish, German, French, or Japanese, the model easily circumvents English-centric defenses and discloses the memorized English PII. Prior research has lacked mechanistic insight into where and how private information propagates and leaks across linguistic boundaries.

To resolve this conflict between multilingual generalization and cross-lingual vulnerability, the authors analyze the internal representations via Logit Lens and latent cosine similarity, revealing a two-stage dynamic: conceptual privacy sharing in the middle layers followed by language-specific projection in the final layers. Core idea: decompose the internal privacy processing pathways into cross-lingual "privacy-universal neurons" and language-exclusive "language-specific privacy neurons" via multilingual integrated gradient attribution, and jointly deactivate both types at inference time to strictly cut the mutual information leakage path while preserving linguistic fluency.

Method

Overall Architecture

The proposed approach begins by constructing MPII, a parallel multilingual PII benchmark across six languages, analyzes the layer-wise information flow via Logit Lens and cross-lingual latent state geometry, and formulates Multilingual Privacy Neuron Control (MPNC). The complete framework unfolds in three sequential stages: tracking cross-lingual information flows to uncover the transition from shared semantic spaces to language-specific outputs, localizing both universal and language-specific privacy neurons through attribution set intersections and differences, and executing coordinated activation deactivation during inference to eliminate the mutual information leakage channel.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multilingual Prompt Input<br/>Q in L"] --> B["Cross-lingual Information Flow Tracing<br/>Logit Lens and latent representation similarity"]
    B --> C["Dual Privacy Neuron Localization<br/>Integrated gradient intersection and difference"]
    C --> D["Multilingual Privacy Neuron Control MPNC<br/>Joint universal and language-specific deactivation"]
    D --> E["Protected Output Generation<br/>Utility-preserved hidden state R'L"]

Key Designs

1. Cross-lingual Information Flow Tracing: Uncovering Two-Stage Internal Privacy Dynamics Existing literature treats privacy memorization as an uninterpretable black box, leaving it unexplained why English-targeted privacy protections fail under other languages. To address this gap, the authors employ Logit Lens to project intermediate hidden representations onto the vocabulary logit space using the unembedding matrix, tracking the Mean Reciprocal Rank (MRR) of sensitive tokens layer by layer alongside cross-lingual cosine similarity. The empirical findings reveal that lower layers retain near-zero MRR for sensitive tokens; across intermediate layers (peaking around layer 14โ€“15 in LLaMA 3.1-8B and layer 22 in Qwen 2.5-7B), cross-lingual cosine similarity surges to 0.7โ€“0.8, marking a shared conceptual space where private information is extracted in a language-agnostic representation. In the final layers (layers 25โ€“32), cross-lingual similarity diverges and drops as representations shift into language-specific vocabulary spaces where leakage risk escalates dramatically. This proves that cross-lingual privacy leakage is governed by a two-stage mechanism of middle-layer conceptual sharing and late-layer language-specific emission.

2. Dual Privacy Neuron Localization: Decoupling Universal and Language-Specific Neurons via Gradient Attribution Standard privacy neuron editing algorithms evaluate attribution solely on English texts, failing to disentangle neurons that govern cross-lingual concept propagation from those that direct language-specific token selection. To overcome this limitation, the method adopts integrated gradient attribution across discrete approximation steps \(m=20\) to quantify the contribution of the \(k\)-th neuron in layer \(l\) with activation \(w^l_k\) to the target PII sequence \(Y\): $\(Att(w^l_k) \approx \beta^l_k \frac{1}{m}\sum_{j=1}^m \frac{\partial P(Y|X, \frac{j}{m}\beta^l_k)}{\partial w^l_k}\)$ Using dual filtering thresholdsโ€”relative attribution peak proportion \(\tau_1=0.1\) and sample occurrence frequency \(\tau_2=0.5\)โ€”candidate privacy neuron sets \(P_\ell\) are constructed for each language subset \(D_\ell\). The framework defines privacy-universal neurons as the intersection across all languages \(P_{uni} = \bigcap_\ell P_\ell\), while language-specific privacy neurons for language \(\ell\) are isolated via set difference \(P_\ell^{(spec)} = P_\ell \setminus P_{uni}\). This decoupling identifies targeted, lightweight neuron subsets comprising only 2.7%โ€“4.5% of the total network parameters.

3. Multilingual Privacy Neuron Control MPNC: Coordinated Suppression under Mutual Information Guarantees Suppression of only universal neurons or only single-language neurons leaves residual leakage paths open. Theoretically, the final representation for target language \(L\) decomposes into \(R_L = (X, U, S_L, N)\), denoting the task-relevant privacy-irrelevant component, universal privacy component, language-specific privacy component, and residual noise, respectively. Theorem 1 establishes that when conditional mutual information terms are strictly positive, deactivating only \(U\) leaves the \(S_L\) channel intact. MPNC intervenes at inference time by clamping both \(U\) and \(S_L\) neuron activations to zero, yielding \(R'_L = (X, 0, 0, N)\) and guaranteeing a strict reduction in representation-level mutual information: $\(I(R_L; P \mid Q, L) - I(R'_L; P \mid Q, L) \ge I(U; P \mid X, Q, L) + I(S_L; P \mid X, U, Q, L) > 0\)$ By the Data Processing Inequality, this directly suppresses privacy leakage in the generated tokens. Causal intervention confirms that deactivating universal neurons yields the primary global leakage reduction, deactivating the query's matching language-specific neurons achieves targeted mitigation, and deactivating unmatching language neurons produces negligible change.

Key Experimental Results

Main Results

Models were fine-tuned for 10 epochs on English MPII records to induce controlled memorization, and then evaluated under multilingual question-answering prompts across six languages (en, zh, es, de, ja, fr) using Mean Reciprocal Rank (MRR, lower is better) and Validation Perplexity (Valid-PPL, lower is better). The table below summarizes the multi-model comparison across all evaluated languages:

Model Method Average MRR (en/zh/es/de/ja/fr) Leakage Reduction vs Orig Average Valid-PPL Generation Fluency Overhead
LLaMA 3.1-8B Original 0.60 โ€” 22.71 Baseline
LLaMA 3.1-8B DEPN 0.53 -11.7% 25.82 +3.11
LLaMA 3.1-8B APNEAP 0.52 -13.3% 24.13 +1.42
LLaMA 3.1-8B MPNC (Ours) 0.46 -23.3% 25.39 +2.68
Qwen 2.5-7B Original 0.62 โ€” 16.92 Baseline
Qwen 2.5-7B DEPN 0.53 -14.5% 18.45 +1.53
Qwen 2.5-7B APNEAP 0.52 -16.1% 17.35 +0.43
Qwen 2.5-7B MPNC (Ours) 0.43 -30.6% 17.67 +0.75
LLaMA 3.2-3B Original 0.62 โ€” 16.23 Baseline
LLaMA 3.2-3B DEPN 0.51 -17.7% 17.55 +1.32
LLaMA 3.2-3B APNEAP 0.52 -16.1% 16.91 +0.68
LLaMA 3.2-3B MPNC (Ours) 0.47 -24.2% 17.76 +1.53

Ablation Study

To verify the distinct causal roles of different neuron categories, the table below reports layer-wise MRR across late transformer layers (Layers 28 to 32) on LLaMA 3.1-8B under various deactivation configurations alongside random baseline suppression:

Deactivation Setting Layer 28 MRR Layer 30 MRR Layer 31 MRR Layer 32 MRR Note
Original (No Intervention) 0.339 0.373 0.520 0.590 Peak leakage risk in final layer
Deactivate Universal Neurons 0.294 0.302 0.428 0.471 Dominant drop across all languages
Deactivate Own Specific Neurons 0.305 0.327 0.484 0.555 Targeted drop for matching prompt language
Deactivate Other Specific Neurons 0.307 0.331 0.492 0.573 Negligible impact on target language
Deactivate Random Neurons 0.297 0.338 0.497 0.580 Ineffective privacy defense with high noise

Key Findings

  • Cross-lingual privacy leakage is pervasive and robust across scripts: despite fine-tuning strictly on English data, email address extraction consistently reaches high MRRs (>0.49) in both Latin (es, fr, de) and non-Latin (zh, ja) languages, with LLaMA 3.2-3B yielding 0.537 MRR under Chinese queries compared to 0.561 in English.
  • Monolingual defenses collapse in multilingual settings: DEPN achieves a 20.8%โ€“23.2% MRR drop under English prompts but only 7.6%โ€“10.0% in Chinese and German; MPNC maintains steady 17.6%โ€“38.5% reductions across all languages.
  • Decisive privacy neurons exhibit severe layer localization: both universal and language-specific privacy neurons cluster in the final 4โ€“6 layers, confirming that while conceptual leakage begins in intermediate layers, actual token emission is decided by late-layer neurons.

Highlights & Insights

  • Uncovers the cross-lingual privacy leakage vulnerability of multilingual LLMs, disproving the prevailing assumption that monolingual privacy interventions suffice for global deployment.
  • Proposes an elegant attribution-based decoupling of universal and language-specific privacy neurons that enables zero-shot inference-time control without costly retraining or parameter updates.
  • Provides rigorous information-theoretic formulation backed by empirical causal interventions, establishing that joint suppression strictly minimizes representation mutual information.

Limitations & Future Work

  • The evaluation relies on fixed question-answer prompt templates; exploring cross-lingual adversarial jailbreaks and multi-turn iterative extraction attacks remains an open avenue.
  • While the MPII benchmark provides parallel texts for names, emails, phone numbers, and job titles, unstructured PII such as physical home addresses and personal URLs are underrepresented and warrant expansion.
  • vs DEPN (Wu et al., 2023): DEPN relies entirely on English attribution to locate and zero out neurons, offering weak protection (~10% reduction) against non-English prompts; MPNC decouples universal from language-specific neurons to provide balanced protection across all languages (23.3%โ€“31.6% reduction).
  • vs APNEAP (Wu et al., 2024): APNEAP applies activation patching on English privacy neurons to mitigate utility degradation; MPNC achieves superior privacy protection with comparably low perplexity increases (+0.75โ€“2.68 PPL).
  • vs Machine Unlearning (DeMem / PME): Fine-tuning-based unlearning and weight editing often induce catastrophic degradation on general multilingual QA and translation benchmarks, whereas MPNC preserves downstream utility on MLQA and FLORES benchmarks.

Rating

  • Novelty: โญโญโญโญโญ First systematic exploration of cross-lingual privacy leakage with an elegant dual-neuron decoupling framework.
  • Experimental Thoroughness: โญโญโญโญโญ Evaluated across three leading open-source models, six diverse languages, causal neuron interventions, and downstream utility tasks.
  • Writing Quality: โญโญโญโญโญ Clear exposition linking mechanistic interpretability insights, formal proofs, and extensive empirical validation.
  • Value: โญโญโญโญโญ High practical relevance for securing multilingual commercial LLM deployments against cross-lingual data extraction.