FlagThis

Abstract

Large Language Models (LLMs) are rapidly moving from isolated research environments to the wild, where they are increasingly deployed on enterprise infrastructure, personal devices, and edge platforms. While cloud-based deployments provide scalable computational resources, they also raise significant security and privacy concerns related to data sovereignty, regulatory compliance, latency, and dependence on third-party providers. As a result, organizations are increasingly adopting edge and on-premise LLM deployments to gain greater control over sensitive data and operational workflows. This architectural shift, however, introduces a new class of security and privacy challenges. In particular, limited computational and memory resources require models to undergo aggressive optimizations, including quantization, pruning, model partitioning, and parameter-efficient adaptation, each of which can introduce new vulnerabilities and fundamentally reshape the threat landscape. We describe this tension as the** _**SecurityEfficiency Paradox**_ **, in which mechanisms designed to improve deployment efficiency may inadvertently weaken model robustness, expose new attack surfaces, or increase privacy risks. In particular, we examine how model compression can affect safety alignment, how partitioned inference can enable information reconstruction attacks, and how continuous local adaptation may introduce privacy leakage and model drift. To systematically analyze these risks, we introduce a deployment-centric taxonomy organized around three fundamental architectural constraints: the Memory Wall, the Quadratic Wall, and the Compute Wall. We then derive a unified constraint model that quantifies when unsafe optimizations become unavoidable, linking each wall directly to specific attack surfaces. Building on this model, we propose the Secure Operational Efficiency Score (SOES), a holistic metric that balances task accuracy, jailbreak resistance, and privacy against energy, memory, and latency, enabling practitioners to select and configure edge LLMs under real-world hardware limits. We further present a practical decision procedure and targeted mitigations for each optimization-induced vulnerability. Together, these contributions move beyond isolated defenses toward a co-designed framework for evaluating security, privacy, and efficiency simultaneously, providing a foundation for securing the next generation of edge-native intelligent systems.