Priviser is a confidential computing platform that enables multiple institutions to jointly compute on combined datasets without moving or exposing the underlying data. Using a proprietary linear-space perturbation method on real numbers, it achieves plaintext-level precision at millisecond speeds, unlike hardware enclaves, homomorphic encryption, or federated learning. The platform unlocks combinatorial insights from previously siloed data, such as clinical and lifestyle cohorts, enabling causal inference and de-biased analysis.
Funding
Funding not disclosed
Founders
Product
Problem
97% of the world's most valuable data has never been used together because datasets cannot move between institutions due to privacy, security, and regulatory constraints. Existing approaches—trusted enclaves, homomorphic encryption, and federated learning—each trade one risk for another, suffering from side-channel exposure, extreme slowdowns, gradient-inversion attacks, or accuracy loss, preventing organizations from unlocking the qualitatively different insights that only combined data can produce.
Solution
Priviser provides a confidential computing platform based on linear-space perturbation on the real number line, a distinct mathematical approach that masks data at the source with random matrices, performs computation on the masks, and aggregates results to reveal the true outcome at plaintext precision. The protocol runs in milliseconds with a constant low-round design, making it compatible with modern AI workloads including CNNs, vision transformers, and transformer attention mechanisms. Unlike hardware or cryptographic alternatives, Priviser delivers results mathematically identical to plaintext computation without trusting silicon vendors or approximating non-linear operations. This enables institutions to combine feature spaces, dissolve confounders, and break selection bias, moving the locus of value creation from individual datasets to the network of combined data.
Target Audience
Primary customers are large institutions in healthcare, finance, and research that hold sensitive datasets and need to jointly compute with other organizations—such as hospitals combining clinical and lifestyle cohorts, or banks merging transactional and behavioral data—without moving or exposing raw data.
Features
- Linear-space perturbation on ℝ with native float64 precision, producing results mathematically identical to plaintext computation
- Constant low-round protocol that completes computations in milliseconds, avoiding the orders-of-magnitude slowdowns of homomorphic encryption
- No reliance on hardware enclaves, eliminating silicon vendor trust and side-channel exposure risks
- Immune to gradient-inversion and membership inference attacks that plague federated learning approaches
- Supports modern AI stack including CNNs, vision transformers, and transformer attention mechanisms
- Enables quadratic and combinatorial interaction spaces across joined datasets, turning observational associations into identified causal effects