Privacy-Preserving Master–Subproblem Decomposition

• The paper introduces decomposition techniques that integrate cryptographic methods to preserve the privacy of agents' local constraints and objective functions during optimization.
• It details methodologies such as alternate projections, cutting-plane oracles, and dual decomposition that maintain privacy while ensuring convergence and feasibility.
• The framework is applied in energy systems, federated learning, and secure multiparty computation, achieving efficient resource allocation and reduced communication overhead.

Privacy-preserving master–subproblem decomposition is a class of distributed optimization methodologies enabling large-scale resource allocation and decision-making while rigorously protecting individual agents’ private information—such as local constraints, objective functions, or datasets—from both centralized operators and peer participants. These frameworks leverage mathematical decomposition (master–subproblem, dual, or task decomposition) and cryptographic or information-theoretic mechanisms to ensure that only aggregated or masked information is revealed at each step. The paradigm is especially prevalent in energy systems, federated learning, secure multiparty computation, and multi-agent operations, where regulatory, commercial, or competitive concerns dictate strict privacy requirements.

1. Formal Problem Models

The privacy-preserving master–subproblem decomposition framework typically starts from a centrally specified optimization problem with both global and agent-specific constraints. Agents possess private feasible sets or objective functions, and the central operator or aggregation node must respect these unknowns in allocating resources or driving consensus.

Resource Allocation with Private Constraints

A canonical model is given by the global problem

$$\min_{p\in\mathbb{R}^T,\,x_1,\dots,x_N\in\mathbb{R}^T} f(p)\qquad \mathrm{s.t.}\quad p\in P,~ x_n\in X_n,~ \sum_{n=1}^N x_n = p,$$

where $P \subset \mathbb{R}^T$ is a convex public constraint set, and each agent $n$ has a private feasible set $X_n$ (Beaude et al., 2019). Only the aggregate variable $p$ is visible to the operator, and the main privacy-preserving challenge is to test whether a given $p$ can be “disaggregated” into private $x_n$ without revealing $X_n$ or $x_n$.

Distributed and Federated Optimization

Decentralized optimization frameworks consider $N$ agents with private cost functions $P \subset \mathbb{R}^T$0, coupled through consensus constraints, expressed as

$P \subset \mathbb{R}^T$1

with threat models targeting both honest-but-curious neighbors and eavesdroppers (Zhang et al., 2018).

Secure Multi-party Computation (SMPC) Decomposition

In SMPC, the global “master” task is a function $P \subset \mathbb{R}^T$2, with each party $P \subset \mathbb{R}^T$3 holding private $P \subset \mathbb{R}^T$4. Decomposition leverages the aggregative structure, splitting into local computations $P \subset \mathbb{R}^T$5 and a smaller SMPC aggregation of these results (Feng et al., 2023).

2. Decomposition Algorithms and Protocol Architecture

The decomposition process introduces an iterative loop involving a master problem (global or coordination function) and subproblems (individual agent computations), with privacy-preserving information flows.

Cutting-plane and Disaggregation Oracles

In resource allocation, the operator maintains a sequence of “cutting-plane” relaxations $P \subset \mathbb{R}^T$6, refining the feasible set with each iteration by introducing hyperplane cuts generated by failed disaggregation attempts. Disaggregation feasibility is tested using the Alternate Projection Method (APM), where projection onto private product sets $P \subset \mathbb{R}^T$7 and constraint hyperplanes is done locally, and feasibility/failure is determined via cycle detection (Beaude et al., 2019).

Dual and Proximal Decomposition

Dual decomposition relaxes coupling constraints via Lagrange multipliers, yielding subproblems per agent that may be solved privately. The master updates dual variables based on residuals of coupling constraints, using subgradient or sophisticated update rules (e.g., Lower-Upper-Bound Switching, LUBS) (Disfani et al., 2015). Proximal Jacobian ADMM variants embed privacy by random function decomposition within the augmented Lagrangian, ensuring agent updates cannot be inverted to reveal private gradients (Zhang et al., 2018).

SMPC Task Decomposition

Aggregative SMPC tasks are split by designing the workflow such that local data transformation precedes a lightweight, cryptographically protected aggregation. Only the result of this aggregation is revealed to participants, ensuring input privacy (Feng et al., 2023).

3. Privacy Preservation Mechanisms

Privacy is guaranteed through both information-theoretic and cryptographic methods specifically tailored to the decomposition framework.

Masking and Secure Aggregation

Aggregate sharing protocols, such as additive secret sharing, enable the computation of sums ($P \subset \mathbb{R}^T$8) necessary for alternate projections or master updates without revealing any individual $P \subset \mathbb{R}^T$9 (Beaude et al., 2019). Each agent splits its vector into random shares, exchanges these with others, and only aggregate results are provided to the operator, satisfying honest-but-curious models for up to $n$0 colluding agents.

Function Decomposition and Obfuscation

Agents split their cost functions into random subcomponents with the property that only the sum is meaningful to the global optimization. The “α-part” interacts with neighbors; the “β-part” is never communicated (Zhang et al., 2018). Observing exchanged variables does not provide enough linear constraints to reconstruct local gradients except at optimum, where KKT conditions necessarily reveal only the optimal point.

Minimal Information Exchange

Dual and LUBS-based protocols restrict communication to coupling constraint values (e.g., aggregate exports/imports, marginal prices), never exposing internal variables such as generator coefficients, network topology, or state variables (Disfani et al., 2015).

SMPC Security Reductions

Task decomposition reduces the input size for the cryptographic primitive to the dimension of the aggregation, not the whole raw data, and preserves full semi-honest SMPC security guarantees (Feng et al., 2023).

4. Theoretical Properties and Convergence

Rigorous convergence and scalability results underpin privacy-preserving decomposition schemes.

Convergence Guarantees

• Alternate Projections: APM converges geometrically to a nearest-point pair in the intersection of private and consensus constraint sets if these are closed and convex. In transportation polyhedra, the convergence rate is linear, with explicit rates depending on problem size (Beaude et al., 2019).
• Cutting-plane Iteration: The outer loop in cutting-plane master–subproblem frameworks requires at most $n$1 cuts (for $n$2 aggregate variables) (Beaude et al., 2019).
• Dual/Subgradient Methods: Dual decomposition converges provided standard convexity, Slater’s condition, and suitable step-sizes; primal iterates converge to optimality (Disfani et al., 2015).
• Proximal Jacobian ADMM: Guarantees $n$3 ergodic convergence under strong convexity and Lipschitz gradient conditions; linear contraction up to a vanishing noise term (Zhang et al., 2018).

Computational Complexity and Scalability

SMPC task decomposition achieves time, memory, and communication costs that are asymptotically constant in dataset size, as all heavy computation (OT, GC) is performed only on the aggregation-level summaries of local transformations (Feng et al., 2023). In cutting-plane frameworks, number of cuts is independent of agent count $n$4, and APM cost scales sublinearly with $n$5 (Beaude et al., 2019).

5. Application Domains and Illustrative Results

Privacy-preserving master–subproblem decomposition frameworks are utilized in power system operations, decentralized control, and privacy-critical machine learning.

Energy Systems

Multi-horizon utility-community interaction is managed through dual and LUBS decomposition, allowing communities and utilities to solve OPF-like tasks with only exchange of aggregate outputs and prices; simulation on a 42-bus feeder with multiple communities demonstrates convergence in 10–15 iterations per hour and practical cost savings (Disfani et al., 2015).

Federated and Decentralized Learning

Agents collaboratively solve consensus or parameter estimation tasks under a privacy model that provably prevents gradient leakage. Simulations on average consensus problems show optimality gap $n$6 is achieved without accuracy loss, outperforming differential privacy alternatives (Zhang et al., 2018).

Secure Machine Learning

SMPC task decomposition is applied to linear regression, logistic regression, and k-means, with the cryptographic bottleneck scaling with the number of parties $n$7 rather than data size $n$8. For example, linear regression time per iteration drops from 300 s (naive SMPC, $n$9) to 50 s post-decomposition, with memory and communication reduced by factors of 4–10 (Feng et al., 2023).

6. Comparative Perspective to Classical Decomposition

Classical Benders and Lagrangian-based schemes are not privacy-preserving, as generating Benders cuts or updating duals involves explicit access to local feasible sets or solution profiles. In contrast, privacy-preserving decomposition leverages either local, functionally obscured procedures (e.g., local APM and Hoffman cuts) or cryptographic protections at the aggregation stage. Lagrangian-based consensus or ADMM methods would otherwise require exchanging state or residuals that reveal private details unless encrypted, while the privacy-preserving frameworks avoid this entirely (Beaude et al., 2019, Zhang et al., 2018, Disfani et al., 2015).

7. Open Questions and Extensions

Future research directions identified include optimal design of decomposition schemes maximizing privacy, extension to time-varying or directed communication graphs, integration with differentially private mechanisms or stronger adversarial models, and application to additional domains with complex privacy constraints (Zhang et al., 2018, Feng et al., 2023).

---

References:

(Beaude et al., 2019, Zhang et al., 2018, Feng et al., 2023, Disfani et al., 2015)