Encrypt Only the Coupling: Private Coordination of Multi-Firm Closed-Loop Supply Networks

A closed-loop supply chain is one where products flow out to customers and then back in for inspection, remanufacturing, or recycling. These networks save the most money and carbon when competing firms share facilities and coordinate flows, but they rarely do so. The reason is that coordination seems to require pooling each firm's freight costs, capacities, and return volumes in one optimizer, and no firm will hand those over. Homomorphic encryption, a way to compute on encrypted numbers without decrypting them, looks like the obvious fix. Run naively, though, it is impractically slow. The closest encrypted decision-maker in the literature needed more than 20 hours for 130 steps of a two-dimensional problem.

This note proposes a different division of labor. Each firm solves its own part of the network design locally, in plaintext, with standard solvers. Only the few quantities that link firms are encrypted: how much each firm sends through a shared facility, and how much shared capacity it uses. Those numbers are summed under encryption, and only the aggregate is ever decrypted. Under this split, the encrypted computation is a sum plus at most one multiplication per coordination round, so no expensive noise-refresh step is ever needed. The research questions become measurable ones. How close does this get to the centralized optimum on cost and carbon? What does it cost per round? And the question we think is under-appreciated: what do the decrypted aggregates leak over many rounds?

Core Claim

In multi-firm forward/reverse logistics networks, firms interact mainly through a small set of linking constraints. These are shared capacity at hybrid facilities that handle both outgoing distribution and incoming inspection, plus the balance of returned volume between collection and recovery. Because of this structure, the encrypted part of a coordinated design can be kept to linear aggregation. We claim this "coupling-only" encryption reaches a cost and carbon gap of a few percent to the centralized plaintext optimum while each round costs milliseconds to seconds, not hours. We also claim that the privacy it delivers depends on how many firms contribute to each linking constraint, and that this dependence can be stated and measured. When only two firms share a facility, an encrypted sum protects nothing, because each firm can subtract its own contribution and recover the other's. We expect this observation to reshape how such systems should be deployed.

What the Prior Work Gives and Leaves Open

The network model comes from A fuzzy robust optimization model for dual objective forward and reverse logistics networks considering carbon emissions. It formulates a five-tier forward/reverse network as a mixed-integer linear program, an optimization over discrete facility choices and continuous flows. The program trades operating cost against carbon emissions under fuzzy and robust uncertainty and is solved centrally with CPLEX. The paper also shows that co-locating distribution and inspection in shared "joint points" yields fixed-cost savings. It supplies our model and our quality reference. It also motivates the gap: its solution assumes a single planner who sees every firm's private parameters. We keep its model and drop that assumption.

Two pieces of encrypted-computation work supply technique and constrain design. Homomorphically Encrypted Linear Contextual Bandit runs a full sequential learner on encrypted data. It matches non-private regret, but approximating matrix inversion, square roots, and argmax as polynomials drives the runtime past 20 hours for $T=130$. We read this as a negative result for the strategy of putting the whole decision procedure inside encryption, and we avoid that strategy. FHAIM reformulates a data-synthesis algorithm so that encrypted steps stay low-degree, for example replacing an $L_1$ norm with a squared $L_2$ norm. It finishes in 11 to 30 minutes on small tables. We reuse its design habit of reshaping the algorithm until the encrypted operations are additions and squares. SecureNets and SAFENet document the accuracy cost of approximating non-polynomial operations and the latency cost of deep encrypted circuits. Our design sidesteps both by keeping every comparison and every max operation in plaintext, either inside a firm or after aggregate decryption.

On the circular-economy side, Optimizing Reserve Decisions in Relief Supply Chains with a Blockchain-Supported Second-Hand E-Commerce Platform shows that shared, tamper-evident records help secondary markets work. It achieves this by making transaction data transparent, which is the opposite of what competing firms want. Our work is complementary: it targets shared computation without shared data. The OARF Benchmark Suite reports that homomorphic encryption in vertical federated learning pushes training toward "hundreds of hours." This is a second warning against encrypting heavy iterative computation wholesale.

Research Program

The coordination method is a decomposition of the joint mixed-integer program. The shared constraints are relaxed with prices, one price per linking constraint. Each firm $i$ then solves its own subproblem over its own variables $x_i$. Firm $i$'s objective is its private cost plus carbon, plus the prices times its use of shared resources. Between rounds, the coordinator updates the prices from the total violation of the shared constraints:

$$\lambda^{k+1} = \max\left(0,\; \lambda^{k} + \alpha_k \Big(\textstyle\sum_i A_i x_i^{k} - b\Big)\right),$$

The symbols are defined as follows. $\lambda^k$ is the vector of prices at round $k$. $A_i x_i^k$ is firm $i$'s use of shared capacity and its contribution to return-flow balance. $b$ is the shared capacity or required balance. $\alpha_k$ is a step size.

Each firm encrypts $A_i x_i^k$ under a threshold version of an approximate-arithmetic homomorphic scheme. In a threshold scheme, decryption requires several parties to cooperate, so no single party can decrypt on its own. An untrusted server adds the ciphertexts. The firms then jointly decrypt only the summed residual. The max operation and the step are applied in plaintext to that aggregate. A second variant adds a quadratic penalty on the residual, which improves convergence. Its encrypted evaluation needs one squared norm, in the FHAIM style, which raises the multiplicative depth to one.

Because the subproblems contain integer facility decisions, prices alone will not always produce a feasible joint plan. In that case we add a final repair step. The firms fix their facility choices from the last round, and the remaining continuous flow problem is coordinated the same way. The fuzzy-robust layer of the original model is kept unchanged. Each firm applies its own conservatism level locally, so the "robustness premium" becomes a per-firm choice rather than a planner's choice.

The analysis has two parts. First, we derive the encrypted operation count, the multiplicative depth, and the ciphertext bytes per round as functions of the number of linking constraints, and check them against measurement. Second, we study leakage. The adversary is an honest-but-curious firm that sees the full sequence of decrypted aggregates and its own contributions, and tries to reconstruct a rival's unit freight cost or return volume. We formalize the fact that with two contributors per constraint, reconstruction is exact. We then measure reconstruction error as the number of contributors grows. Finally, we test one mitigation: adding calibrated noise to the decrypted aggregate, so that noise is added once per round to the outputs rather than to every firm's inputs.

Baselines

The naive reference point is no coordination. Each firm optimizes its own tier of the network alone, with no shared joint-point facilities. This is the status quo the proposal must beat on cost and carbon. No reported result was found on Lacuna for this uncoordinated configuration, so we compute it ourselves.

The quality ceiling is the centralized plaintext solution of Li et al. on the "Ws company" automotive remanufacturing network. That network has 10 customer nodes, 5 distribution centers, 5 recovery facilities, 3 plants, and 3 disposal sites. The paper reports total cost of 1,769,715 yuan and emissions of 29,480,588 g CO2 in the certain scenario, and 1,882,260 yuan in the most conservative scenario. We rebuild the instance from the parameters the paper reports. Where parameters are missing, we generate instances at the same scale and re-solve them centrally with the same solver family, so that all comparisons share data and solver budgets. Beating this baseline is not the goal. Coming within a few percent of it on both objectives counts as success.

The strongest no-cryptography competitor is the same decomposition run in plaintext, with each firm's shared-resource usage revealed to the coordinator. It reaches exactly the same plans as our method. The comparison therefore isolates two things: the overhead the encryption adds, and the information it hides.

The strongest privacy-preserving competitor for the aggregation step is additive secret sharing among the firms. We run it on the same rounds, the same network latency model, and the same number of parties, and add simulated firm dropouts.

The "encrypt everything" reference is a design that evaluates the price update and the feasibility repair entirely under encryption, including comparisons. For this we have only indirect evidence: HELBA's more than 20 hours for 130 steps in Garcelon et al.. We will estimate its cost by operation counting rather than claim a run.

Input perturbation, where each firm adds noise to its shared quantities before revealing them in plaintext, is run at matched leakage levels.

Two ablations isolate the key ingredients. The first compares the depth-zero linear price update with the depth-one penalized variant, to show what the extra multiplication buys in convergence. The second runs the method with and without output noise, to separate the convergence benefit of encryption from its privacy benefit.

Why not just…?

Why not just run the decomposition in plaintext and have firms reveal their shared-facility usage? This is the strongest simple competitor, because it produces identical plans at zero cryptographic cost. What it gives up is that every round reveals each firm's volume through each shared facility. Over dozens of rounds, those reveals expose the firm's response curve to price, and from that its private costs. Our leakage experiment measures exactly this. If a rival can recover costs from the plaintext trajectory but cannot from aggregates when three or more firms share a constraint, encryption has earned its place. If the two leak equally, the cryptography is not justified, and we will say so.

Why not just use secret sharing among the firms instead of homomorphic encryption? For pure sums this is a real threat, because secret-sharing aggregation is cheap. The argument for encryption is operational. Recyclers and logistics providers connect intermittently, and encryption lets an untrusted server aggregate asynchronously without every party staying online for every round. FHAIM makes the same argument against multi-round protocols. We will measure wall-clock time per round and completion under dropouts for both methods. If secret sharing wins on both, the contribution narrows to the decomposition and the leakage analysis, which remain valid under either primitive.

Why not just add differential-privacy noise to each firm's shared numbers? Noise added to physical quantities such as freight capacity or return volume pushes the joint plan toward infeasibility in tightly constrained networks. Garcelon et al. show that noise on inputs to a sequential learner degrades regret from roughly $\sqrt{T}$ to $T^{3/4}$. Our matched-leakage comparison tests directly whether noise added once to the decrypted aggregate costs less optimality than noise added to every firm's inputs.

Why not just encrypt the whole optimization? Branching, comparisons, and argmax under encryption need high-degree polynomials and repeated noise refreshing. This is precisely why HELBA took more than 20 hours on a tiny problem, and why SAFENet works hard to remove non-polynomial operations. Our operation-count estimate quantifies the gap. We do not expect this approach to be competitive.

Decisive Checks

The proposal succeeds if four checks pass. First, the coupling-only method is within a few percent of the centralized optimum on both cost and carbon on the Ws-scale network and on larger synthetic networks, and clearly beats the uncoordinated baseline. Second, each round requires no noise refresh, with measured cost per round orders of magnitude below the encrypt-everything estimate; we will also test GPU arithmetic via cuHE. Third, cost reconstruction by a curious rival is exact with two contributors per linking constraint and degrades measurably as contributors are added, while plaintext decomposition stays fully leaky. Fourth, output-side noise reaches a given leakage level at lower optimality loss than input noise.

The outcome that would most change our view is secret sharing matching encryption under dropouts while leakage is unchanged. In that case we would recast the work as a study of which multi-firm circular networks can be coordinated privately at all, with the choice of cryptographic tool as a secondary matter.