Secure Multi-User Linearly-Separable Distributed Computing
Amir Masoud Jafarpisheh, Ali Khalesi, Petros Elia
TL;DR
This work addresses secure multi-user linearly-separable distributed computing by casting user requests as a matrix factorization $\mathbf{F}=\mathbf{D}\mathbf{E}$ and introducing information-theoretic secrecy guarantees. It shows two fundamental criteria: (i) per-user secrecy requires the visible common randomness to span a subspace of dimension $\alpha_k-1$, and (ii) removing the observed servers from $\mathbf{D}$ leaves a rank of at least $K-1$, yielding a universal secrecy condition across fields. A cost-preserving secure transformation augments $\mathbf{E}$ with a basis for $\mathrm{Null}(\mathbf{D})$ and injects shared randomness, achieving perfect secrecy over GF$(q)$ and controllable leakage over $\mathbb{R}$ while preserving correctness and costs. The framework also provides a converse bound on the communication cost $\delta \le 1 - \frac{K-1}{N}$ and demonstrates concrete scheme realizations, including a detailed example and a general design, with implications for secure distributed analytics in federated and multi-tenant settings.
Abstract
The introduction of the new multi-user linearly-separable distributed computing framework, has recently revealed how a parallel treatment of users can yield large parallelization gains with relatively low computation and communication costs. These gains stem from a new approach that converts the computing problem into a sparse matrix factorization problem; a matrix $F$ that describes the users' requests, is decomposed as \(F = DE\), where a \(γ\)-sparse \(E\) defines the task allocation across $N$ servers, and a \(δ\)-sparse \(D\) defines the connectivity between \(N\) servers and \(K\) users as well as the decoding process. While this approach provides near-optimal performance, its linear nature has raised data secrecy concerns. We here adopt an information-theoretic secrecy framework, seeking guarantees that each user can learn nothing more than its own requested function. In this context, our main result provides two necessary and sufficient secrecy criteria; (i) for each user \(k\) who observes $α_k$ server responses, the common randomness visible to that user must span a subspace of dimension exactly $α_k-1$, and (ii) for each user, removing from \(\mathbf{D}\) the columns corresponding to the servers it observes must leave a matrix of rank at least \(K-1\). With these conditions in place, we design a general scheme -- that applies to finite and non-finite fields alike -- which is based on appending to \(\mathbf{E}\) a basis of \(\mathrm{Null}(\mathbf{D})\) and by carefully injecting shared randomness. In many cases, this entails no additional costs. The scheme, while maintaining performance, guarantees perfect information-theoretic secrecy in the case of finite fields, while in the real case, the conditions yield an explicit mutual-information bound that can be made arbitrarily small by increasing the variance of Gaussian common randomness.