Privacy-Preserving Computation
Session also emphasizes privacy-preserving computation, particularly in the training and deployment of AI models. Privacy-preserving computation is essential for safeguarding sensitive data during processing, especially in AI model training and deployment.
Technologies like Secure Multi-Party Computation (MPC) and homomorphic encryption enable computations on encrypted data, ensuring privacy throughout the process.
Secure Multi-Party Computation (MPC):
MPC allows multiple parties to jointly compute a function over their inputs while keeping those inputs private. This means that each party's data remains confidential, and only the final computation result is revealed. Traditional data handling often involves decrypting data for computation, which exposes it to potential breaches. MPC addresses this by enabling computations directly on encrypted data, maintaining confidentiality throughout.
Integrating with Extended MPC Protocols:
Nillion has developed a novel MPC protocol that extends the capabilities of Linear Secret-Sharing Schemes (LSSS) to perform certain non-linear operations, specifically evaluating sums of products of non-zero user inputs. This protocol operates in two phases:
Pre-processing Phase: This phase prepares the network to securely handle high-value data, generating and distributing shares (masks) for each factor and term in the sum of products expression using standard MPC techniques. Importantly, this pre-processing is independent of the actual input values and depends only on the number of inputs (factors and terms).
Non-Interactive Computation Phase: In this phase, the actual computations on the private inputs are performed. It involves three stages:
Input Stage: Shares generated during pre-processing are distributed to parties, who combine them with their inputs to create masked factors. These masked factors are broadcasted to the network, maintaining information-theoretic security.
Evaluation Stage: Parties perform local calculations on the masked factors, leveraging their multiplicative homomorphic properties to enable operations like addition and multiplication directly on the masked data.
Output Stage: Parties reveal the results of their local calculations, which are then aggregated to derive the final computation result.
This protocol offers several advantages:
Non-linear Arithmetic Capabilities: It can evaluate sums of products of hidden inputs without leaking information to any party.
Efficient Pre-processing: The creation of shares is independent of input values and depends solely on the number of inputs in each term.
Asynchronous Computation: The non-interactive nature of the computation phase aligns with asynchronous workflows and accelerates the process, as it does not require message exchanges between parties.
Information-Theoretic Security (ITS): The protocol upholds the ITS inherent in LSSS, meaning it is secure against adversaries with unlimited computing resources and time.
By integrating such advanced MPC protocols, Session can ensure that sensitive data remains confidential during computation, enhancing privacy and security in AI model training and deployment.
Last updated