Privacy-Preserving Artificial Intelligence: Challenges, Techniques, and Trade-offs

Abstract

Modern artificial intelligence (AI) depends on large and diverse datasets, creating significat challenges for the protection of personal and sensitive information. The privacy problem is not limited to model traning: data can be exposed during collection, preparation, training, deployment, inference, and subsequent use of AI outputs. This paper discusses the privacy challenges introduced by AI systems and surveys several privacy-preserving techniques, including data minimization, anonymization, and differential privacy, that can reduce the exposure of sensitive personal data while maintaining the utility of AI models.