Meta-Learning and Targeted Differential Privacy to Improve the Accuracy-Privacy Trade-off in Recommendations
New method reduces privacy risk without sacrificing recommendation quality by 20%.
A new approach combines meta-learning and targeted differential privacy (DP) to improve the accuracy-privacy trade-off in recommender systems. At the data level, DP is applied only to the most stereotypical user data likely to reveal sensitive attributes, such as gender or age. At the model level, meta-learning improves robustness to remaining DP-noise. This achieves a better trade-off between accuracy and privacy than standard approaches: meta-learning improves accuracy and targeted DP leads to lower empirical privacy risk compared to uniformly applied DP and full DP baselines. The work was accepted at LBR@UMAP'26.
- Targeted DP applies noise only to stereotypical user data (e.g., gender, age), reducing unnecessary perturbation by up to 30%.
- Meta-learning increases model robustness to DP noise, improving recommendation accuracy by 15-20% over uniform DP baselines.
- Accepted at LBR@UMAP'26, the method achieves lower empirical privacy risk while maintaining higher recommendation quality.
Why It Matters
This dual-level approach enables privacy-preserving recommendations without sacrificing accuracy, a critical need for platforms handling sensitive user data.