FL + Secure Aggregation
Goal
Train a shared model while hiding individual participant updates from the coordinator.
Actors
Participants, coordinator, model owner, secure-aggregation service, auditors, and downstream model users.
Data Flow
flowchart LR
I[Participant registry] -->|identity + eligibility| C[Coordinator]
K[Key setup] -->|round keys| A
K -->|round keys| B
K -->|round keys| D
C[Coordinator] -->|global model + training code| A[Site A]
C -->|global model + training code| B[Site B]
C -->|global model + training code| D[Site C]
A -->|masked update| S[Secure aggregation]
B -->|masked update| S
D -->|masked update| S
S -->|aggregate update| C
C -->|new global model| U[Model users]
C -->|round logs + metrics| L[Audit log]
S -->|dropout + threshold events| L
Trust Boundaries
| Boundary | What crosses | Who can see it | Risk |
|---|---|---|---|
| Site to coordinator | Training code, model version, round instructions | Site operators, coordinator | Bad code or wrong model version |
| Site to aggregation | Masked updates and metadata | Aggregation service | Update leakage if threshold/key assumptions fail |
| Aggregation to coordinator | Aggregate update | Coordinator | Small rounds can expose participants |
| Coordinator to users | Final model | Model users | Memorization and membership inference |
| System to logs | Metrics, errors, round metadata | Operators, auditors | Logs can reveal participant behavior |
Assumptions
- Enough participants complete each round to satisfy secure-aggregation thresholds.
- Participant identity and key setup are reliable.
- Local training code is reviewed and versioned.
- The coordinator cannot inspect individual unmasked updates.
Assumption Review
| Assumption | How to validate | If it fails |
|---|---|---|
| Round threshold is large enough | Simulate dropouts and small-site participation | Aggregate updates can expose a site or force skipped rounds |
| Participant identity is reliable | Bind sites to keys, certificates, and round eligibility | A fake or compromised participant can poison or observe protocol behavior |
| Training code is controlled | Version code, configs, and model hashes per round | Sites may train incompatible or malicious updates |
| Logs are minimized | Review round metadata, errors, and support bundles | Metadata can reveal participation, timing, or site behavior |
PET Stack
Federated learning, secure aggregation, participant authentication, optional DP, robust aggregation, and model auditing.
Common PET Combinations
| Add | Use when | New risk |
|---|---|---|
| Differential privacy | The released model needs record-level or patient-level contribution bounds | Utility loss and accounting complexity |
| Robust aggregation | Participants may be malicious or compromised | Harder to combine with hidden individual updates |
| TEEs for orchestration | Coordinator code should be constrained by attestation | Hardware trust and side-channel assumptions |
| Output review | Final model or metrics may leak membership | Requires release gates and model-audit ownership |
What This Does Not Protect Against
- Poisoned updates by malicious participants.
- Leakage from the final model.
- Weak local security at participant sites.
- Small-round inference.
- Debug logs that expose update metadata.
Out of scope unless explicitly added: malicious-secure training, compromised local data systems, model extraction by downstream users, and side-channel leakage from participant infrastructure.
Deployment Notes
Plan for participant dropouts, versioned training code, reproducible evaluation, secure key setup, and rollback when aggregation fails.
Tradeoffs
Secure aggregation improves update privacy but makes debugging, anomaly detection, and malicious-client handling harder.
Failure Modes
Gradient leakage without aggregation, poisoning, small participant rounds, key setup errors, weak participant identity, and plaintext operational logs.
Evaluation Checklist
- What minimum number of participants is required per round?
- Are update sizes, timing, and dropouts logged safely?
- Is the final model tested for memorization?
- Can poisoning be detected without inspecting individual updates?
- Is the key setup recoverable after participant failure?
References
- Bonawitz et al., Practical Secure Aggregation for Privacy-Preserving Machine Learning on User-Held Data, 2016/2017.
- Zhu, Liu, and Han, Deep Leakage from Gradients, NeurIPS 2019.