MPC Analytics Pipeline
Goal
Compute a joint analytic result across parties without revealing each party's raw inputs to a central processor.
Actors
Data-contributing parties, protocol participants, query requester, output reviewer, auditors, and downstream decision makers.
Data Flow
flowchart LR
Q[Query requester] -->|approved query| P[Policy review]
P -->|schema + protocol spec| V[Schema validator]
V -->|validated spec| A[Party A]
V -->|validated spec| B[Party B]
A -->|secret shares / protocol messages| M[MPC protocol]
B -->|secret shares / protocol messages| M
M -->|allowed output| O[Output reviewer]
O -->|released metric| U[Decision makers]
M -->|run metadata| L[Audit log]
O -->|release decision| L
Trust Boundaries
| Boundary | What crosses | Who can see it | Risk |
|---|---|---|---|
| Requester to review | Query and intended output | Reviewers | Query may be too revealing |
| Parties to protocol | Shares and protocol messages | Protocol parties | Collusion or malformed inputs |
| Protocol to reviewer | Computed output | Reviewer | Output leakage |
| Reviewer to users | Released metric | Decision makers | Misuse or overinterpretation |
| Protocol to audit log | Run metadata and release decisions | Auditors | Logs can reveal participation or query patterns |
Assumptions
- Collusion threshold is explicit.
- Parties authenticate each other and run the agreed protocol.
- Output policy is defined before computation.
- Malformed inputs are validated or handled.
Assumption Review
| Assumption | How to validate | If it fails |
|---|---|---|
| Collusion threshold is realistic | Compare protocol threshold to ownership, hosting, and incentives | Inputs can be reconstructed by parties treated as separate |
| Output policy is enforceable | Review examples of allowed, suppressed, and rejected outputs | MPC computes a private input function whose result still leaks |
| Parties are available | Test retries, timeouts, and participant dropout behavior | Latency and aborts can make the workflow unusable or revealing |
| Schemas match | Validate definitions, units, identifiers, and missing values | The result may be wrong even if the protocol is secure |
PET Stack
MPC, participant authentication, schema validation, query approval, output thresholding, optional DP, and audit logging.
Common PET Combinations
| Add | Use when | New risk |
|---|---|---|
| Differential privacy | Aggregate output can reveal small cohorts or repeated-query differences | Utility loss and budget accounting |
| PSI | The workflow starts with entity overlap | The match set may be sensitive |
| Clean-room workflow | Analysts need governed query submission and review | Platform trust and policy bypasses |
| TEEs | Protocol coordination or preprocessing needs confidential execution | Hardware trust and attestation |
What This Does Not Protect Against
- Outputs that reveal sensitive facts.
- Collusion beyond the stated threshold.
- Malicious inputs if the protocol is only semi-honest.
- Poor schema alignment.
- Operational metadata leakage.
Out of scope unless explicitly added: malicious security, denial of service by a party, side channels in protocol implementations, and misuse of released metrics.
Deployment Notes
Estimate rounds, bandwidth, availability requirements, and failure behavior before committing to the protocol.
Tradeoffs
MPC reduces reliance on one trusted processor but increases protocol, networking, debugging, and participant-coordination complexity.
Failure Modes
Unrealistic collusion assumptions, high latency, participant unavailability, malformed inputs, tiny-cohort outputs, and opaque cost.
Evaluation Checklist
- What collusion threshold is claimed?
- Is malicious security required?
- What outputs are allowed or suppressed?
- How are schemas validated?
- What happens when a party drops out?