Package: homomorpheR 1.0

Introduction | The global problem and its consensus split | The federated fit in the clear | The same fit under threshold FHE | Comparison with the centralized fit | What the protocol hides and reveals | Discussion | Limitations

The statistical problem | The privacy constraint | The protocol | Implementation | 1. Generate a CKKS key pair | 2. Build workers and master | 3. Run mle() through the encrypted channel | Beyond MLE | CAVEAT

The statistical problem | DLBCL lymphoma cohort | The aggregated fit | The protocol | Wiring up the protocol | Iterative MLE through the encrypted protocol | Comparison with the cleartext fit | What just happened | What this demonstrates | Caveats and extensions

Introduction | The setup | Threat model | The protocol | Synthetic data | Threshold key generation | Joint rotation keys | Per-site adapter fit and database setup | Diagonal-encoded matrix-vector multiply | Inner product against the local database | What is in the slots we do not read | Site contribution_fn: full per-site protocol step | Master orchestration and threshold-decrypted top-k | Reference-truth comparison | What the adapter buys: fidelity, posture, and a sweet spot | Discussion | Limitations | Related work | References

Introduction | The problem and its consensus split | Adding noise | A brief output-DP primer | The protocol modification | The stopping rule changes | Setup | Simulated cohort | Choosing $\rho$ and $T$ without touching the cohort | Threshold-FHE setup | The DP-ADMM loop | Centralized CVXR fit | The $\sigma$ sweep | Privacy budget | References

The problem this package addresses | Terms you will meet | How the pieces fit together | Sites that are not in this R session | What "the right answer" means | Which vignette to read | A note on Paillier | References

Introduction | A brief output-DP primer | The protocol | The Cox setup (same DLBCL data as the other Cox vignettes) | Threshold setup and DP-noised workers | Mechanical correctness: $\sigma = 0$ reproduces cox-threshold | Increasing noise: where does it break? | Nelder–Mead at the same noise scales | Privacy budget | References

Introduction | Threat model | The Cox setup (same DLBCL data as cox.Rmd) | The protocol | Implementation | Iterative MLE through the threshold protocol | Comparison with the cleartext fit | Discussion | Limitations

The problem | The target answer | Why the classic solution needed two non-cooperating parties | Threshold encryption removes the need for masking | The threshold-BFV implementation | The encrypted answer is exact | Summary

Introduction | Exact, for integer schemes | Approximate, for CKKS

The use case | Scenario: weighted biomarker score | Setup: hospital's context and patient data | Lab side: apply the model to encrypted data | Hospital side: decrypt the results | Verification | Transport guarantees | Threat model: model extraction from the hospital side | Layered defenses | Network protocol: serialization

Overview | The setting | Step 1: train in the clear | Step 2: encrypt patient data | Step 3: evaluate the linear predictor (encrypted) | Step 4: apply the sigmoid (encrypted) | Step 5: decrypt predictions | Summary | Limitations

The problem | Setup | The aggregator sets up encryption | Each site computes locally and encrypts | The aggregator aggregates | Verification | The details of what happened | Serializing for a real distributed protocol