A Light-Weight Neural Network Model for Privacy Preserving Recommendation System
Biswas Rudra Jyoti Arka, Hafiz Imtiaz · Measurement Interdisciplinary Research and Perspectives · 2026
Recommendation systems rely heavily on user data, analyzing which can have the potential to leak personal and sensitive information about a person posing significant privacy and security risks. While differential privacy offers a rigorous framework to mitigate these threats, integrating it into the training of neural network-based recommendation systems often degrades performance due to induced randomness – originating the concept of privacy-utility trade-off. Conventional matrix factorization-based recommendation methods frequently fail to balance these conflicting needs effectively. To address this, we propose an artificial neural network (ANN)-based collaborative filtering model designed for a superior privacy-utility trade-off. We analyze the privacy loss of model training using Rényi differential privacy and evaluated the model on real datasets against various privacy and dataset parameters. Finally, we compare the results with existing non-private and differentially private algorithms, and show that our ANN-based collaborative filtering model can guarantee strong privacy with an excellent privacy-utility trade-off.