Secure Multiparty Computation(MPC)

Secure Multiparty Computation (MPC) is a field of cryptography aimed at creating methods for parties to jointly compute a function over their inputs while keeping those inputs private. This technology enables different entities, such as organizations or individuals, to collaborate and perform computations without exposing their private data to each other. The essence of MPC is to ensure that no party learns anything more than the output of the computation.

The concept of MPC was first introduced in the 1980s and has evolved significantly since then. It addresses a variety of practical scenarios where sensitive data must be protected. For instance, companies could use MPC to compare customer data for overlaps without revealing the actual data to each other, or government agencies could collaborate on crime or fraud detection without compromising on data confidentiality.

MPC works by breaking data into fragments and distributing these fragments among the computing parties. Each party performs computations on their fragment and shares only specific results. The full result is only revealed when all fragments are combined in a particular way. Advanced cryptographic techniques, including homomorphic encryption and zero-knowledge proofs, are often employed to facilitate these operations.

Using MPC for AI model inference allows for the utilization of sensitive data in making powerful machine learning predictions, adhering to high privacy standards. The process begins with the sharing of a trained AI model. This model can either be pre-trained by one party and then shared as encrypted fragments among all participants, or it could be collaboratively trained with all parties contributing their data, which remains encrypted throughout the training process. Each participant then encrypts their input data using cryptographic techniques that are compatible with MPC protocols, ensuring the data can be processed without revealing any sensitive information.

Inference is carried out through distributed computation, where each party processes parts of the encrypted model with their encrypted data. This is facilitated by MPC protocols such as SPDZ or SecureNN, which outline how to securely share, process, and combine data without compromising privacy.

After processing, the intermediate results are securely aggregated. This step ensures that no single party can access the entire dataset or the intermediate computational states, maintaining the confidentiality of the data. The final step involves the collaborative decryption of the result, where only the final output—such as the inference from the AI model—is revealed.