An Empirical Evaluation of Differential Privacy in Wide & Deep Recommender Systems

Abstract

Recommender systems are central to modern digital platforms and rely on rich user interaction data to deliver personalized content. However, such data-driven models raise privacy concerns, especially when deployed at scale. Recommender systems can be viewed as large-scale behavioral data analysis pipelines, where machine learning models extract latent preference structures from high-dimensional interaction data. Differential privacy (DP) provides formal guarantees against individual data leakage but remains underexplored in deep hybrid recommender architectures. This study evaluates the integration of DP into a Wide & Deep recommendation model using differentially private stochastic gradient descent (DP-SGD). We compare private and non-private training regimes on the MovieLens 1M dataset and analyze the resulting privacy-utility trade-off. Model performance is evaluated using RMSE and MAE, while privacy loss is quantified through Rényi differential privacy. Our findings show that the model maintains competitive predictive accuracy even under strong privacy guarantees (e.g., σ = 2.0, ε = 0.24), with stable behavior at higher noise levels. The results provide empirical evidence that privacy-preserving deep models can maintain analytical utility while securing sensitive behavioral data, contributing to the development of privacy-preserving data-driven systems.