How mainak9093/privacy-preserving-recommendation works
OblivRec combines private matrix-factorization training with private recommendation delivery. Training factors ratings under replicated 2-of-3 sharing; at serving time, a user reconstructs scores and selects items locally, then fetches each record with DPF-PIR. The graph also shows the cleartext modeling/evaluation path, PIR-query harvesting, and the MPC and FSS building blocks. Some subsystem links are based on README descriptions or the SOURCE INDEX; unsampled implementation details are not inferred.
Main entry points: demo.cpp Recommendation demo · pir.cpp DPF-PIR · factor.cpp Private factorization · mpc.cpp Replicated MPC
Components 24
Product workflow
- Training runner
src/apps/train.cpp - Recommendation demo
src/apps/demo.cpp - Serving server
src/apps/server.cpp - Shared-score serving
src/serve/serve.cpp
Private training
- Private factorization
src/mf/factor.cpp - Cleartext factorization
model/mf.py
Private delivery
- DPF-PIR
src/pir/pir.cpp - DPF keys
src/dpf/dpf.cpp - Fixed-width catalogue
src/pir/catalogue.cpp - Query harvesting
src/pir/harvest.cpp - TCP framing
src/net/tcp.cpp
Cryptographic substrate
- Replicated MPC
src/mpc/mpc.cpp - Shared matrix-vector
src/mpc/mvp.cpp - Truncation and normalization
include/oblivrec/nonlinear.hpp - FSS comparison gate
src/mpc/nonlinear/msnzb.cpp - Replicated shares
include/oblivrec/share.hpp - Pseudorandom generator
src/common/aes.cpp - System randomness
src/common/csprng.cpp
Modeling and analysis
- Ratings data
model/data.py - Recommendation metrics
model/metrics.py - Fetch-observer attack
model/attack.py - Privacy study
model/dp_study.py - Poisoning study
model/poison_study.py
Other components
- User
Connections 28
- User → Recommendation demo: requests recommendations
README.md - Training runner → Private factorization: runs private training
src/apps/train.cpp - Private factorization → Replicated MPC: computes shared products
- Private factorization → Truncation and normalization: uses nonlinear protocols
- Private factorization → Shared matrix-vector: uses matrix-vector work
- Replicated MPC → Replicated shares: exchanges replicated shares
- Truncation and normalization → Replicated MPC: uses MPC operations
include/oblivrec/nonlinear.hpp - Truncation and normalization → FSS comparison gate: uses comparison gate
src/mpc/nonlinear/msnzb.cpp - Recommendation demo → Shared-score serving: scores shared embedding
src/apps/demo.cpp - Recommendation demo → DPF-PIR: fetches selected records
src/apps/demo.cpp - Recommendation demo → TCP framing: sends network frames
src/apps/demo.cpp - Serving server → DPF-PIR: answers PIR keys
src/apps/server.cpp - Serving server → Fixed-width catalogue: loads catalogue
src/apps/server.cpp - DPF-PIR → DPF keys: evaluates query keys
- DPF-PIR → Fixed-width catalogue: reads records
- DPF keys → Pseudorandom generator: expands pseudorandom tree
- DPF keys → System randomness: generates key material
src/dpf/dpf.cpp - Query harvesting → DPF-PIR: uses PIR queries
- Query harvesting → Replicated MPC: shares harvested data
- Cleartext factorization → Ratings data: loads rating splits
model/mf.py - Cleartext factorization → Recommendation metrics: evaluates predictions
model/mf.py - Fetch-observer attack → Ratings data: uses dataset
model/attack.py - Fetch-observer attack → Cleartext factorization: fits baseline factors
model/attack.py - Privacy study → Fetch-observer attack: uses attack analysis
model/dp_study.py - Privacy study → Recommendation metrics: measures quality
model/dp_study.py - Poisoning study → Ratings data: uses dataset
model/poison_study.py - Poisoning study → Recommendation metrics: measures quality
model/poison_study.py - Training runner → Recommendation demo: provides trained model
README.md