Multi-Party Computation (MPC)

Banks can jointly compute a loan applicant's credit score without sharing their proprietary risk models or the applicant's private financial data. Each bank holds a segment of the applicant's financial history. Using MPC, they compute a combined risk assessment, revealing only the final score. This enables more accurate lending decisions while maintaining strict compliance with financial privacy regulations like GDPR and CCPA.

• Key Mechanism: Secure comparison and aggregation of encrypted risk factors.
• Privacy Guarantee: No single bank learns another's model weights or the client's full financial profile.
• Industry Impact: Reduces default risk through collaborative intelligence without data pooling.

Hospitals and research institutions can identify genetic markers for diseases by performing statistical analysis on combined patient genomes. MPC allows computation of p-values and correlation coefficients across the joint dataset. No institution reveals its patients' genomic sequences. This accelerates medical discovery for conditions like cancer or rare diseases while preserving patient confidentiality and institutional IP.

• Core Function: Statistical computations (e.g., linear regression, chi-square tests) on encrypted genomic data.
• Regulatory Alignment: Enables HIPAA-compliant and GDPR-compliant cross-border research collaborations.
• Example: The iDASH competition has featured MPC challenges for secure genome analysis.

Advertisers, publishers, and analytics firms can measure campaign performance—such as reach, frequency, and attribution—without sharing user-level data. MPC protocols allow the matching of encrypted user identifiers across datasets to compute aggregate metrics. This enables performance tracking in a post-third-party-cookie world, aligning with Apple's App Tracking Transparency and Google's Privacy Sandbox initiatives.

• Core Challenge: Securely joining datasets on a common key (e.g., a hashed user ID) without revealing the key.
• Output: Aggregate reports (e.g., 'Campaign X reached 1M unique users') with formal privacy guarantees.
• Industry Shift: Moves analytics from data aggregation to computation aggregation.

When two organizations hold different features about the same entities (e.g., a bank has credit history, and an e-commerce site has purchase history), they can collaboratively train a machine learning model. MPC secures the process of gradient computation and parameter updates during training. The resulting model, which leverages both feature sets, is more powerful than any model trained on a single source, yet neither party sees the other's raw feature data.

• Architecture: Often combined with Homomorphic Encryption for core linear algebra operations.
• Distinction from Federated Learning: FL typically assumes partitioned samples; MPC here addresses partitioned features.
• Use Case: Joint anti-money laundering models between a bank and a fintech platform.

A decentralized machine learning paradigm where a global model is trained collaboratively across multiple edge devices or servers holding local data samples, without exchanging the raw data itself. Clients compute updates on local data and send only model parameters (e.g., gradients) to a central server for aggregation.

• Key Difference from MPC: FL focuses on collaborative training while keeping data decentralized; MPC is a general cryptographic protocol for secure computation of any function on private inputs. FL often uses MPC techniques for Secure Aggregation.

A form of encryption that allows specific types of computations to be performed directly on ciphertext. The encrypted result, when decrypted, matches the result of operations performed on the plaintext.

• Contrast with MPC: HE enables computation on encrypted data by a single party. MPC involves multiple parties jointly computing a function where each party holds a secret share of the data. HE is often more computationally intensive, while MPC distributes the computational and trust burden.

A rigorous mathematical framework for quantifying and limiting privacy loss. It guarantees that the inclusion or exclusion of any single individual's data in a dataset has a negligible effect on the output of a computation.

• Synergy with MPC: DP and MPC are complementary. MPC can protect data during computation, while DP can be applied to the final output to prevent inference attacks on the results. Together, they provide strong end-to-end privacy guarantees.

A specific cryptographic protocol, often built using MPC, used in Federated Learning. It allows a central server to compute the sum of client model updates without being able to inspect any individual client's contribution.

• Mechanism: Each client encrypts or secret-shares its update. The server aggregates the ciphertexts/shares and only learns the final aggregated model update, not the individual vectors. This prevents the server from reverse-engineering a client's private data from their gradient update.

Cryptographic protocols that allow one party (the prover) to prove to another party (the verifier) that a statement is true without revealing any information beyond the validity of the statement itself.

• Application with MPC: ZKPs can be integrated into MPC protocols to allow parties to prove they are following the protocol correctly (e.g., that their local computation was performed honestly on valid inputs) without revealing their private inputs, enhancing security against malicious participants.