Security Challenges in Large Language Models Across the Data Life Cycle: A Cryptography-Aware Comprehensive Review

Abstract

Large Language Models (LLMs) are now embedded in security and privacy critical applications, yet they remain vulnerable to attacks that span their entire data life cycle. This survey provides a comprehensive, cryptography-aware review of these risks across three phases—training, inference, and deployment; while explicitly connecting them to classical security goals and primitives. We introduce a simple stage-wise risk scoring model inspired by NIST risk assessment that propagates vulnerabilities across the life cycle, and we instantiate it with a numeric example linking training time poisoning to inference time data extraction. We further propose a life cycle aligned evaluation framework that maps modern benchmarks (e.g., HarmBench, JailbreakBench, TrustLLM, DecodingTrust) to concrete threat classes and reports representative quantitative results, such as attack success rates under different defenses. Finally, we analyze the practicality of advanced defenses—including differential privacy, fully homomorphic encryption, secure multi-party computation, and zero knowledge proofs—in light of their computational overhead and deployment constraints, building on foundational cryptography and privacy works. Our goal is to bridge the gap between classical cryptographic theory and emerging LLM specific threats, and to outline research directions toward secure, privacy preserving, and rigorously evaluated LLM pipelines.