Overview
Secure Multi-Party Computation (SMPC) enables multiple parties to jointly compute a function over their combined inputs while keeping each party's individual input private. No party learns anything beyond the output of the computation and what can be inferred from their own input and the output.
SMPC supports GDPR Article 5(1)(c) data minimization by eliminating the need to centralize data, Article 25(1) data protection by design by building privacy into the computation architecture, and Article 26 joint controller arrangements by enabling collaborative analytics without data sharing.
Core SMPC Techniques
Shamir Secret Sharing
Shamir's Secret Sharing (1979) splits a secret value into n shares such that any t shares can reconstruct the secret (threshold t-out-of-n), but fewer than t shares reveal no information.
Properties:
- Information-theoretic security (unbreakable regardless of computational power)
- Supports addition of shared values without communication
- Multiplication requires an interactive protocol (Beaver triples or resharing)
Application: Splitting personal data across multiple servers such that no single server (or coalition below threshold) can reconstruct the data.
Garbled Circuits
Yao's Garbled Circuits (1986) enable two-party computation. One party (the garbler) encrypts a boolean circuit; the other party (the evaluator) evaluates the encrypted circuit without learning intermediate values.
Properties:
- Constant-round protocol (efficient for high-latency networks)
- Communication cost proportional to circuit size
- Best for two-party computation with complex boolean functions
Oblivious Transfer
Oblivious Transfer (OT) is a protocol where a sender has multiple messages and a receiver selects one message to receive, without the sender learning which message was selected and without the receiver learning the other messages.
Role in SMPC: OT is the foundational building block for garbled circuit evaluation and is used for input wire labels in Yao's protocol.
SMPC Frameworks
| Framework | Protocol | Language | Security Model | Best For |
|---|---|---|---|---|
| MP-SPDZ | Multiple (SPDZ, MASCOT, semi-honest, malicious) | Python-like DSL | Semi-honest and malicious | Research and prototyping with multiple security models |
| CrypTen | Secret sharing (2-party and 3-party) | Python (PyTorch) | Semi-honest | ML inference and training on shared data |
| MOTION | GMW, BMR, arithmetic/boolean sharing | C++ | Semi-honest | High-performance 2+ party computation |
| ABY/ABY3 | Arithmetic, Boolean, Yao sharing | C++ | Semi-honest (2-party ABY, 3-party ABY3) | Mixed-protocol computation (switching between share types) |
| Sharemind | Additive secret sharing (3-party) | SecreC (DSL) | Semi-honest (honest majority) | Enterprise deployment with managed infrastructure |
MP-SPDZ Protocol Selection
| Protocol | Security Model | Parties | Preprocessing | Online Performance |
|---|---|---|---|---|
| SPDZ/MASCOT | Malicious (dishonest majority) | 2+ | Heavy (OT-based) | Fast online phase |
| Semi2k | Semi-honest (dishonest majority) | 2+ | Moderate | Fast |
| Shamir | Semi-honest (honest majority) | 3+ | Light | Very fast |
| Rep3 | Semi-honest (honest majority) | 3 | None | Very fast |
| MASCOT | Malicious (dishonest majority) | 2+ | OT-based | Moderate |
| Yao's GC | Semi-honest | 2 | Garbling | Fast evaluation |
Architecture for GDPR-Compliant SMPC
┌─────────────────────────────────────────────────────────────┐
│ SMPC Computation Coordinator │
│ (orchestrates protocol execution) │
└──────┬─────────────────┬─────────────────┬──────────────────┘
│ │ │
┌──────▼──────┐ ┌───────▼──────┐ ┌──────▼──────┐
│ Party A │ │ Party B │ │ Party C │
│ (Bank) │ │ (Insurer) │ │ (Retailer) │
│ │ │ │ │ │
│ Input: x_A │ │ Input: x_B │ │ Input: x_C │
│ Share: [x_A]│ │ Share: [x_B] │ │ Share: [x_C]│
│ │ │ │ │ │
│ Compute on │ │ Compute on │ │ Compute on │
│ local share │ │ local share │ │ local share │
└──────┬──────┘ └───────┬──────┘ └──────┬──────┘
│ │ │
└────────────────►│◄────────────────┘
│
┌──────▼──────┐
│ Output: │
│ f(x_A,x_B, │
│ x_C) │
│ (revealed │
│ to all) │
└─────────────┘
Each party learns ONLY the final output f(x_A, x_B, x_C),
NOT the individual inputs of other parties.
Practical Deployment Patterns
Pattern 1: Private Set Intersection (PSI)
Two parties determine which records they have in common without revealing records unique to either party.
Use case: A bank and an insurer identify shared customers for a joint risk assessment without revealing their full customer lists.
| Step | Action | Privacy Guarantee |
|---|---|---|
| 1 | Each party hashes their customer identifiers with a shared key | Individual identifiers not revealed |
| 2 | Parties engage in PSI protocol (DH-based or OT-based) | Only intersection elements revealed |
| 3 | For matched records, proceed with joint computation on shared data | Non-matched records remain private |
Pattern 2: Private Aggregation
Multiple parties compute aggregate statistics (sum, mean, count) over their combined data without revealing individual contributions.
Use case: Three regional offices of Prism Data Systems AG compute total headcount and average salary without revealing per-office figures.
Pattern 3: Private ML Training
Multiple parties jointly train a machine learning model using SMPC to protect training data during the process.
Use case: Using CrypTen, two hospitals jointly train a logistic regression model on encrypted patient features.
Performance Considerations
| Factor | Impact | Mitigation |
|---|---|---|
| Network latency | SMPC is communication-intensive; high latency increases total time | Co-locate computation nodes; use constant-round protocols (Yao) for WAN |
| Data size | Communication scales with input size | Pre-aggregate locally where possible; use sketching |
| Circuit complexity | Deeper circuits require more communication rounds (for GMW) | Optimize circuit depth; use Yao for constant rounds |
| Number of parties | More parties = more communication pairs | Use 3-party protocols with honest majority for efficiency |
| Security model | Malicious security is 10-100x slower than semi-honest | Use semi-honest for trusted consortium members; reserve malicious for adversarial settings |
Implementation Workflow
- Define the computation — Express the joint function as arithmetic or boolean operations.
- Select security model — Semi-honest if all parties are trusted (e.g., consortium members); malicious if adversarial behavior is possible.
- Select framework — Choose MP-SPDZ for flexibility, CrypTen for ML workloads, Sharemind for managed enterprise.
- Implement protocol — Write the computation in the framework's DSL, test on synthetic data.
- Benchmark — Measure end-to-end latency, communication volume, and memory usage.
- Deploy — Establish secure communication channels (mTLS), key distribution, and monitoring.
- Govern — Execute joint controller agreement (Art. 26), update Article 30 records, document in DPIA.
Key Regulatory References
- GDPR Article 5(1)(c) — Data minimization (no data centralization needed)
- GDPR Article 25(1) — Data protection by design
- GDPR Article 26 — Joint controller arrangements for SMPC consortia
- GDPR Article 28 — Processor obligations (SMPC coordinator role)
- GDPR Article 32(1) — Security measures
- GDPR Recital 78 — Technical measures for data protection
- ENISA Report: Data Protection Engineering (2022)