Research map / Federated learning and privacy / Communication-efficient FL
SecEmb: Sparsity-Aware Secure Federated Learning of On-Device Recommender System with Large Embedding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SecEmb, a lossless secure federated recommender protocol that uses function secret sharing to privately retrieve and aggregate only the sparse rated-item embedding updates, cutting user communication and computation costs.
Problem
Federated recommender systems transmit full embedding models and updates, which is costly for edge devices; existing sparsity-aware protocols leak rated item indices or update coordinates to the server.
Approach
Encode each user's rated item embedding retrieval and sparse row-wise embedding gradient aggregation as point functions under two-server function secret sharing, and reuse the FSS evaluation path across the retrieval and aggregation stages so that servers learn only the aggregated model.
Key results
- SecEmb reduces upload communication by up to 91x (MF on Yelp; reduction ratio 91.22) relative to secure FedRec with two-server ASS.
- User-side computation time is reduced by up to about 70x on Yelp with MF and FM, per the paper.
- Against lossy compression (Bit8Quant, TernQuant, SVD, CoLR), SecEmb keeps accuracy: compression degrades RMSE on average by 1.29% to 2.04%, while SecEmb is lossless.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SecEmb instead of secure FedRec | MF · Yelp | upload communication cost: 47.81 -> 0.52 MB; 91.22 | Helped |
| Use SecEmb instead of lossy compression | MF · ML1M | RMSE: $0.916$ $\pm 0.001$ -> $\mathbf{0.903}$ $\mathbf{\pm 0.002}$ | Helped |
| Use SecEmb instead of lossy compression | Caser · ML1M | NDCG@10: 0.263 -> 0.270 | Helped |
| Use initial construction without row-wise encoding or shared retrieval path (ablation) | SecEmb, MF · ML100K | upload communication cost: 0.17 -> 4.56 MB | Hurt |
| Use row-wise encoding but not shared retrieval path (ablation) | SecEmb, MF · ML100K | upload communication cost: 0.17 -> 0.28 MB | Hurt |
| Use initial construction without row-wise encoding or shared retrieval path (ablation) | SecEmb, MF · ML1M | computation cost: 0.47 -> 15.35 milliseconds | Hurt |
| Use row-wise encoding but not shared retrieval path (ablation) | SecEmb, MF · ML1M | computation cost: 0.47 -> 0.90 milliseconds | Hurt |
| Use SecEmb instead of lossy message compression | MF · Across evaluated datasets | prediction accuracy: 1.29%, 1.74%, 1.95%, and 2.04% | Helped |
Surprising findings
- For DeepFM on ML100K and ML1M (item sizes below 4k), SecEmb gives only about 1.05x to 1.24x upload reduction because dense parameters dominate.
Limitations
- Efficiency advantage shrinks when the per-user number of rated items m' approaches m or when dense parameters dominate (e.g., DeepFM on small datasets).
- Requires two non-colluding servers.
- Differential privacy integration is only analyzed empirically on ML1M, with noticeable utility loss at small epsilon.
Experimental setup
Datasets: MovieLens 100K MovieLens 1M MovieLens 10M MovieLens 25M Yelp Amazon
Benchmarks: MovieLens 100K MovieLens 1M MovieLens 10M MovieLens 25M Yelp ML1M Amazon
Metrics: upload communication cost (MB) download communication cost user computation time (ms) RMSE reduction ratio HR@10 NDCG@10
Baselines: Secure FedRec with two-server ASS SecAgg Bit8Quant TernQuant SVD CoLR DP-FedRec SecEmb-Init SecEmb-RowEnc
Builds on: function secret sharing secure aggregation two-server additive secret sharing differential privacy (Gaussian mechanism, moments accountant)
Closest papers by meaning
- Rethinking Byzantine Robustness in Federated Recommendation from Sparse Aggregation Perspective (AAAI 2025 · similarity 0.87)
- Federated Recommendation with Additive Personalization (ICLR 2024 · similarity 0.84)
- RFLPA: A Robust Federated Learning Framework against Poisoning Attacks with Secure Aggregation (NeurIPS 2024 · similarity 0.83)
- Towards Trustworthy Federated Learning with Untrusted Participants (ICML 2025 · similarity 0.83)
- Federated Recommendation with Explicitly Encoding Item Bias (AAAI 2025 · similarity 0.82)
- Federated Graph Learning for Cross-Domain Recommendation (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.