Privacy-preserving in foundation models: a systematic review of techniques, threats, and trade-offs

Abstract Foundation Models (FMs) are large-scale Artificial Intelligence (AI) models that have been trained on vast amounts of data. These models have gained great attention in the field of AI due to their evolving capabilities and their potential to transform various domains. However, such opportunities come with a wide range of privacy and security challenges along the lifecycle of the FMs including the leakage of sensitive training data or the exposure of models and users’ input. This systematic literature review analyzes the evidence from 295 peer-reviewed studies published from 2022 to 2025. The study focuses on privacy-preserving techniques, what they are, where they apply in the FM lifecycle, what threats they address or mitigate, their effectiveness, and main challenges. The study also analyzes privacy threats, their prevalence in FMs, and the main challenges to address them. Then we conduct a deep analysis of the privacy-utility trade-offs addressed in the literature, how they are formulated, optimized, and evaluated. The review provides a lifecycle-aware taxonomy for privacy-preserving techniques and privacy threats, including a deep look at trends and gaps related to privacy-utility trade-off formulation and measurement. The aim is to guide researchers, professionals, and policy makers in designing AI FMs that are robust, private, and ethical.