Cross-Border Data Security and Privacy Risks in Large Language Models and IoT Systems

Chalitha Handapangoda · 2026

Read Cross-Border Data Security and Privacy Risks in Large Language Models and IoT Systems by Chalitha Handapangoda (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-06612.txt.

Introduction

Chalitha Handapangoda tackles cross-border data risks for large language models and connected devices in a 2026 arXiv paper. When data moves between countries, it falls under clashing privacy laws. Examples include Europe's General Data Protection Regulation and China's Personal Information Protection Law. Technical flaws, such as a model memorizing private text, add to the danger.

Handapangoda proposes an architecture, meaning a planned system design, that builds privacy in from the start. It adjusts encryption to local rules, adds adaptive noise to protect individuals, and checks compliance in real time using cryptographic proofs. In a simulated multi-country test, unauthorized exposure fell below five percent with no compliance violations. Model usefulness stayed above ninety percent.

Catalog record

CREATORChalitha Handapangoda
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-06612

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Cross-Border Data Security and Privacy Risks in Large Language Models and IoT Systems Chalitha Handapangoda Abstract The reliance of Large Language Models and Internet of Things systems on massive, globally distributed data flows creates systemic security and privacy challenges. When data traverses borders, it becomes subject to conflicting legal regimes, such as the EU’s General Data Protection Regulation and China’s Personal Information Protection Law, compounded by technical vulnerabilities like model memorization. Current static encryption and data localization methods are fragmented and reactive, failing to provide adequate, policy-aligned safeguards. This research proposes a Jurisdiction-Aware, Privacy-by-Design architecture that dynamically integrates localized encryption, adaptive differential privacy, and real-time compliance assertion via cryptographic proofs. Empirical validation in a multi-jurisdictional simulation demonstrates this architecture reduced unauthorized data exposure to below five percent and achieved zero compliance violations. These security gains were realized while maintaining model utility retention above ninety percent and limiting computational overhead. This establishes that proactive, integrated controls are feasible for secure and globally compliant AI deployment. I Introduction The rapid expansion of AI infrastructure has fundamentally changed how sensitive data is generated, processed, and transmitted. Both Large Language Models (LLMs) and Internet of Things (IoT) systems require massive, globally distributed data flows that inevitably cross legal and organizational boundaries. This essential cross-border movement introduces complex security and privacy challenges because data leaves the original controller’s legal and technical zone of protection [ 9 ] . The resultant fragmentation in global governance means that data becomes subject to irreconcilable legal frameworks, such as the GDPR, the U.S. CLOUD Act, and data localization requirements in China [ 6 ] . These international legal and technical inconsistencies significantly complicate compliance and governance, leading to uneven enforcement of data protection standards. Addressing this global data flow dilemma is both interesting and important, as the integrity of global AI deployment relies on resolving these jurisdictional conflicts. I-A Technical Vulnerabilities and Gaps Beyond the legal quandaries, critical technical vulnerabilities persist. Prior work confirms that LLMs can memorize sensitive training data and unintentionally reveal this private information during inference, exposing user data even without direct access to the underlying storage layer [ 2 ] . Similarly, IoT devices continuously transmit sensitive personal and environmental data (e.g., location, health metrics) to cloud regions worldwide, creating additional cross-border attack surfaces [ 10 ] . This problem is hard because naive approaches fail due to this technical-legal interdependence. Static encryption schemes, the simplest protective measure, are vulnerable to mandatory decryption laws. Data localization prevents geographic movement but fails entirely to address inference-time leakage from a model already deployed globally. Current protection methods are insufficient because they treat security and compliance as separate, reactive concerns, rather than integrated, proactive design requirements following frameworks like the NIST Privacy Framework [ 7 ] or ISO/IEC 27701 [ 5 ] . I-B Proposed Solution and Contribution This study addresses these systemic risks by introducing the Jurisdiction-Aware, Privacy-by-Design (JAPD) architecture . The core principle of the JAPD approach is the proactive, unified treatment of legal compliance and technical security. This integrated architecture dynamically applies localized encryption policies, enforces Differential Privacy (DP) mechanisms at both the training and inference stages, and asserts real-time compliance validation through cryptographic audit proofs. The primary contribution is demonstrating the technical feasibility of achieving high regulatory compliance and robust security guarantees simultaneously, without imposing unacceptable computational costs. The remainder of this paper is structured as follows. Section II reviews related research, comparing our method to prior works in DP theory and adversarial machine learning. Section III presents a motivating example, detailing specific cross-border threat scenarios using the STRIDE framework. Section IV presents the core hypothesis and the technical implementation of the JAPD architecture. Section V presents the empirical evidence, validating the architecture’s performance, utility, and compliance effectiveness against the NIST Privacy Framework. Finally, Section VII summarizes the conclusions and outlines key directions for future work. II Related Research The global deployment of intelligent systems has created complex cross-border data flows that challenge both technical security and legal compliance. The goal of this section is to compare the novel JAPD architecture against foundational and contemporary works, demonstrating our advancement. II-A Inference-Time Vulnerabilities (Carlini et al., 2021) Carlini et al. identified a critical privacy risk posed by LLM memorization, demonstrating that sensitive training data can be extracted through carefully crafted queries [ 2 ] . Their work established that model weights act as sensitive data reservoirs, which conventional Differential Privacy (DP), formalized by Dwork and Roth [ 3 ] and applied to deep learning by Abadi et al. [ 1 ] , cannot fully protect once the model is globally deployed. The JAPD architecture builds upon this finding by moving the solution to the model inference stage. We introduce integrated inference-time DP that dynamically adjusts the privacy budget based on the user’s jurisdiction, providing real-time, context-aware protection that the original training-only DP schemes could not offer. II-B Conflicting Data Sovereignty Mandates (Khan, 2025) Khan systematically reviewed the fragmentation of global privacy laws, emphasizing the irreconcilable legal differences between regimes such as the EU’s GDPR and China’s PIPL [ 6 ] . The core problem identified is the lack of harmonized, legally enforceable standards for cross-border data transfer, a requirement underscored by compliance standards like ISO/IEC 27701 [ 5 ] . Our method differs fundamentally from this prior work by technically operationalizing this legal challenge. JAPD employs Dynamic Jurisdiction-Aware Routing and cryptographic audit proofs, transforming abstract legal mandates into concrete, enforceable network policies. II-C Limitations of Distributed Learning (Zhang et al., 2022) Zhang et al. demonstrated the efficacy of Federated Learning (FL) for maintaining data locality in IoT systems, but acknowledged that FL provides uniform privacy guarantees that are jurisdiction-agnostic [ 10 ] . This creates a use case that FL cannot handle: scenarios where heterogeneous regulatory constraints require adaptive privacy controls (e.g., stricter protection for EU data). The JAPD architecture advances upon FL by introducing Localized Key Management and Multi-Layer Encryption, ensuring cryptographic keys remain bound to their legal jurisdictions—a necessary step for key sovereignty that FL ignores. III Motivating Example: Cross-Border Threat Scenarios The critical need for a jurisdiction-aware architecture is underscored by specific cross-border threat scenarios, analyzed using the standard STRIDE framework [ 8 , 4 ] . The core danger is the rapid expansion of trust boundaries, where each jurisdiction introduces new intermediaries and inconsistent controls that adversaries can exploit. III-A Scenario 1: LLM Model Memorization and Cross-Border Privacy Leakage This scenario involves LLMs trained on globally distributed datasets that inadvertently memorize personal or confidential information. • Information Disclosure: The primary threat is extraction attacks revealing Personally Identifiable Information (PII). When a model trained using EU data (protected by GDPR) is deployed in a jurisdiction with weaker laws, the stored model weights become a conduit for exposing sensitive data that is no longer legally protected [ 2 ] . • Repudiation: Multi-jurisdictional deployments obscure accountability. Inadequate audit trails across borders hinder the ability to definitively prove the origin of the leakage or assign responsibility. III-B Scenario 2: IoT Data Interception During Cross-Jurisdictional Transit IoT devices continuously transmit highly sensitive data, such as behavioral patterns and health metrics, to remote cloud centers, often traversing numerous national borders. • Information Disclosure: As data packets

Read the complete text