Differential Privacy in Machine Learning: A Survey from Symbolic AI to LLMs

Researchers Francisco Aguilera-Martínez and Fernando Berzal have released a comprehensive academic survey, 'Differential Privacy in Machine Learning: A Survey from Symbolic AI to LLMs,' providing a crucial roadmap for securing AI development. The 236-page work, published on arXiv (ID: 2506.11687), systematically reviews Differential Privacy (DP), a mathematical framework that ensures an algorithm's output doesn't reveal whether any single individual's data was used in its training. This is foundational for preventing models like GPT-4 or Claude from memorizing and leaking sensitive personal information.

The survey is notable for its historical scope, tracing DP's application from early Symbolic AI systems through to today's large language models (LLMs). It goes beyond theory to analyze practical proposals and methods for integrating DP during model training, and crucially, describes how these privacy-preserving techniques can be evaluated in real-world scenarios. By consolidating years of research into one accessible document, the authors aim to accelerate the adoption of robust privacy safeguards, which is increasingly critical as companies train models on vast, potentially sensitive datasets including emails, code, and medical records.

This work arrives at a pivotal moment for the AI industry, where regulatory pressure and public concern over data privacy are intensifying. It serves as an essential reference for engineers and researchers building the next generation of AI, providing the technical foundation needed to develop systems that are both powerful and trustworthy. The survey effectively bridges the gap between abstract cryptographic theory and the practical demands of modern machine learning pipelines.