When several organizations jointly train a model or run a shared optimization — power grids, ridesharing fleets, federated learning — a popular way to protect each party's data is to avoid any central coordinator. Instead, neighboring nodes solve local subproblems and exchange only small "boundary" messages: prices, dual variables, consensus copies of a few shared quantities. The privacy argument is architectural: since nobody ever sees raw data or full objective functions, nothing private can leak. In a parallel literature, adversarial machine learning has shown the opposite happens with gradients: because a neural network is a deterministic differentiable machine, a shared gradient is a structural blueprint of the input that produced it, and modern attacks reconstruct that input almost perfectly. These two communities have barely spoken to each other. This proposal argues that the boundary messages of decentralized optimization are more exposed than a single gradient, because an eavesdropping neighbor observes not one snapshot but an entire trajectory — many correlated, low-dimensional vectors produced by a known deterministic recursion — and that cumulative observation can defeat the "under-determined, therefore private" argument on which the strongest decentralized methods rest.
The plan has three parts. First, build the attack: extend optimization-based inversion from single gradients to primal–dual trajectories and run it against operator-splitting protocols that have never faced an empirical inversion attack. Second, quantify the leakage: connect it to the recently derived topological privacy bounds so we can predict, before running anything, which network layouts are safe. Third, build the fix: a decaying, pairwise, zero-sum perturbation on the exchanged boundary messages that cancels within any group of honest nodes, survives imperfect consensus, and preserves the exact optimum.
Core Claim
In an operator-splitting method such as ADMM, gradient tracking, or primal-dual decomposition, node $i$ sends its neighbor a message $u_i^k$ at iteration $k$ that is a deterministic function of the node's private data, its current local state, and its dual variables: $u_i^k = g(x_i, y_i, \lambda_i^k)$. An honest-but-curious neighbor knows $g$ and observes the whole sequence $\{u_i^k\}_{k=1}^{K}$. The core claim is that per-message under-determination does not survive aggregation across iterations: if the boundary message has dimension $d_b$ and the private unknowns (say, a batch of $n$ training samples) have dimension $n > d_b$, then a single message leaves $n - d_b$ degrees of freedom unresolved, but $K$ messages give the adversary roughly $K \cdot d_b$ constraints, and the trajectory becomes over-determined once $K \gtrsim n / d_b$. The adversary's problem is exactly gradient matching lifted to trajectories:
$$\arg\min_{\hat{x},\,\hat{y}} \; \sum_{k=1}^{K} \Big(1 - \frac{\langle g(\hat{x},\hat{y},\hat{\lambda}^k),\, u_i^k\rangle}{\|g(\hat{x},\hat{y},\hat{\lambda}^k)\|\,\|u_i^k\|}\Big) + \alpha\, \mathrm{TV}(\hat{x}),$$
where the cosine-distance term and the total-variation image regularizer follow Inverting Gradients - How easy is it to break privacy in federated learning?, which showed that gradient direction carries the semantic content of an input while gradient magnitude mostly reflects training progress. The dual sequence $\hat{\lambda}^k$ is inferred jointly, since duals follow the same known recursion the adversary can replay. Our claim has a defensive mirror image: if the masking noise applied to messages is pairwise on communication edges, zero-sum in aggregate, and decays to zero over iterations, then any set of honest nodes cancels it internally, an adversary observing one edge sees only noise it cannot separate from the message, and the consensus error left by imperfect cancellation is a bounded, vanishing tail sum rather than a permanent bias.
What Prior Work Does and Does Not Cover
Four papers sit closest to this seam, and each leaves it open in a different way. FedNew: A Communication-Efficient and Privacy-Preserving Newton-Type Method for Federated Learning is the strongest operator-splitting privacy claim in the file: clients replace transmitted gradients and Hessians with a composite ADMM vector $y_i^k$ whose mapping from private quantities is provably under-determined per round. But its privacy argument is purely analytical — it never runs an inversion attack against $y_i^k$, and its guarantees assume convex objectives. The inversion literature has already shown that under-determination is not a wall: Gradient Inversion with Generative Image Prior breaks it with generative priors, Inverting Gradients breaks it with total-variation regularization, and High-Fidelity Gradient Inversion in Distributed Learning breaks it for batches up to size 64 with stepwise layer-by-layer optimization. Our attack applies exactly this machinery to ADMM vectors and boundary trajectories, which nobody has done.
On the attack side, On the (In)security of Peer-to-Peer Decentralized Machine Learning proved that decentralized learning is less private than federated learning: gossip updates are dominated by one-hop neighbors, and its "functional marginalization" attack isolates a single victim's contribution. But it attacks plain stochastic gossip and local model averaging — not operator-splitting methods with dual variables, and not the power-system boundary exchanges it would be natural to target. Similarly, Topology-Dependent Privacy Bound For Decentralized Federated Learning proves that an adversary can at best reconstruct the aggregate over its "honest component" — the connected subgraph of non-colluding nodes — with per-node leakage scaling as roughly $1/(2(m-1))$ for component size $m$, and validates the bound with gradient inversion that succeeds completely when $m = 1$. But its analysis assumes plain average consensus; it explicitly excludes dual variables and primal–dual dynamics, which is precisely the regime our proposal analyzes. Finally, Private and Accurate Decentralized Optimization via Encrypted and Structured Functional Perturbation (EFPSN) is the closest existing defense: neighbors exchange encrypted noise offline so that perturbed cost functions sum to the true global objective, and the noise cancels at consensus. It is our main structural inspiration, but it perturbs cost functions rather than message trajectories, assumes perfect consensus, and was validated only against the early iDLG attack on MNIST digits. Our defense differs on all three counts, and our attack will first test whether EFPSN's masking even survives modern directional and generative inversion.
Research Program
The attack work (phase one) formalizes the trajectory-inversion objective above and instantiates it in three variants of increasing strength: Euclidean matching following Deep Leakage from Gradients, cosine matching with a total-variation prior following Inverting Gradients, and a prior-guided variant restricting the search to a generative image manifold following Gradient Inversion with Generative Image Prior. A key technical question is how to handle the dual recursion: the adversary must replay the neighbor's ADMM or gradient-tracking updates, which we solve by making the dummy duals decision variables of the same outer optimization, initialized from the observed sequence. We will attack four protocols: plain decentralized gradient tracking in the style of Optimal gradient tracking for decentralized optimization as the reference, the FedNew single-step-ADMM protocol, EFPSN-masked decentralized optimization, and a microgrid-style boundary exchange where neighbors see only voltage magnitudes and phases or trade volumes and clearing prices — the setting motivating distributed frameworks like A Distributed Operational Method for Convex Hull Pricing Based on the Alternating Direction Method of Multipliers with Dantzig–Wolfe and Benders Decomposition and A Scalable, Distribution Network-Aware, Customer Privacy-Preserving Framework for Operation of Virtual Power Plants, where the target of reconstruction is the private load or cost profile rather than an image.
The theory work (phase two) extends the honest-component bound of Topology-Dependent Privacy Bound from average consensus to primal–dual updates. The interesting wrinkle is that dual variables converge to persistent shadow prices — they do not average away like stochastic gradients — so we expect leakage through dual trajectories to decay more slowly in component size $m$ than the $1/(2(m-1))$ consensus bound, and part of the program is to derive the correct rate. We also connect to Privacy Amplification by Iteration for ADMM with (Strongly) Convex Objective Functions, which shows iteration itself can amplify privacy for convex objectives; our trajectory attack is the natural empirical stress test of whether that amplification holds once the adversary sees all intermediate duals rather than only the limit.
The defense (phase three) perturbs each exchanged message $u_i^k$ with noise $\Delta_i^k$ constructed pairwise on communication edges, so that for every honest connected component $C$ in the topology,
$$\sum_{i \in C} \Delta_i^k = 0 \quad \text{for all } k, \qquad \|\Delta_i^k\| \leq \beta_k, \quad \sum_k \beta_k < \infty,$$
with $\beta_k$ a decaying schedule. Edge-pairwise construction means cancellation holds for any honest subgraph, not just the full network, so colluding hubs cannot fragment the cancellation the way they fragment average-consensus privacy in the power-law topologies studied in Topology-Dependent Privacy Bound. The decaying schedule means the residual consensus bias under finite step sizes, packet loss, or early stopping is bounded by the tail sum $\sum_{k' \geq k} \beta_{k'}$ and vanishes at convergence — directly repairing the perfect-consensus assumption that EFPSN acknowledges as its main limitation. Unlike EFPSN we do not perturb cost functions, so the method composes with any existing operator-splitting solver, and unlike FedNew we do not rely on under-determination, whose failure our own attack will have demonstrated or bounded.
Baselines
The "do nothing" reference point is plain decentralized training with no defense: nodes run gradient tracking or ADMM consensus and exchange raw updates, and we attack them with the original Euclidean DLG procedure of Deep Leakage from Gradients. This gives the leakage ceiling; for calibration, the mixed-quantization paper reports that DLG on undefended CIFAR-10 gradients reaches a PSNR of 55.3, which is near-perfect reconstruction, and Inverting Gradients reports roughly 14 dB PSNR with cosine matching on a trained ResNet-20 versus roughly 7 dB for Euclidean matching. Our trajectory attack must at minimum match these numbers on the same data and architectures to establish that it is a competent attack before we use it to probe supposedly private protocols.
The strongest existing method at this intersection is EFPSN, Private and Accurate Decentralized Optimization via Encrypted and Structured Functional Perturbation, which reports maintaining high accuracy on MNIST logistic regression up to noise magnitude $\gamma = 10^4$ (where non-zero-sum perturbation collapses around $\gamma = 10^0$) and a privacy budget at least four orders of magnitude larger than non-zero-sum baselines. We will run EFPSN exactly as specified, on its own benchmark (MNIST, LeNet-5 and logistic regression) and on ours, under two conditions it never faced: modern directional and generative inversion, and imperfect consensus (asynchronous updates, 5–20% message loss, early stopping). Beating EFPSN means our defense keeps reconstruction fidelity at or below the level EFPSN achieves under the weak iDLG attack while EFPSN's resistance fails under the modern attacks, and keeps the optimality gap small where EFPSN's residual noise bias grows.
The second strongest competitor is FedNew, whose per-round under-determination is the leading analytical privacy argument for operator splitting; it reports nearly 10x fewer transmitted bits than standard FedNew on the LibSVM w8a dataset with quantization, but no reported result found on Lacuna for any empirical inversion attack against its transmitted vectors. We run its protocol unchanged on its own benchmarks (a1a, w7a, w8a, phishing) plus CIFAR-10, and beat it by showing either that trajectory or prior-guided inversion reconstructs its clients' data (falsifying the implicit claim), or, if it genuinely holds, by characterizing the iteration count $K$ at which it breaks. Local defenses — Gaussian differential-privacy noise, gradient pruning, a decentralized port of the CENSOR orthogonal-subspace replacement of CENSOR: Defense Against Gradient Inversion via Orthogonal Subspace Bayesian Sampling, and mixed quantization in the style of Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks — are run as defense baselines under identical data, topologies, and tuning budgets (same grid-search range over their noise/sparsity/quantization parameters, same communication allowance). Finally, two ablations isolate our ingredients: edge-pairwise noise without the decaying schedule (tests whether decay is what buys robustness to imperfect consensus), and globally-canceling noise in EFPSN's style applied to messages rather than functions (tests whether per-component cancellation is what buys robustness to colluding hubs).
Why not just…?
Why not just add differential-privacy noise to the boundary messages? This is the simplest defense and must be run, and it is a real threat at small network sizes. But the file's evidence is consistently negative at privacy-useful noise levels: Deep Leakage from Gradients reports that Gaussian or Laplacian noise stops reconstruction only at variance above $10^{-2}$, where accuracy collapses; Mixed Quantization reports DP-driven accuracy of roughly 78–85% on MNIST where quantization holds 96.67%; and On the (In)security of Peer-to-Peer Decentralized Machine Learning notes local DP in decentralized settings is strictly worse than central DP. Our defense is best understood as DP-style noise whose sum is zero over honest components, which is why it can be large per-message without biasing the optimum. The head-to-head experiment — same privacy target, same tuning budget — settles it, and if plain DP matches our accuracy at equal reconstruction resistance we will report that honestly.
Why not just use homomorphic encryption or secure aggregation end-to-end? Cryptography does eliminate inversion in principle, but the cost evidence in the file is decisive for the decentralized regime we target: Mixed Quantization characterizes homomorphic encryption as computationally prohibitive for edge devices with tedious multi-client key generation, and On the (In)security of Peer-to-Peer Decentralized Machine Learning notes secure aggregation carries heavy overhead and is hard to deploy on peer-to-peer topologies where every node aggregates from many neighbors. EFPSN itself restricts Paillier encryption to a one-time offline phase for exactly this reason. Our defense deliberately avoids any per-iteration cryptography; the only cryptographic-adjacent step is a one-time pairwise noise exchange, comparable to EFPSN's offline phase and far lighter than encrypting every message.
Why not just deploy CENSOR or mixed quantization locally at each node? These are the strongest cheap local defenses: CENSOR reports roughly 86% CIFAR-10 accuracy — matching the undefended baseline — at 0.00236% computational overhead, and mixed quantization drops DLG reconstruction on CIFAR-10 from PSNR 55.3 to 8.1. But both were designed for client-server federated learning, and the file is explicit that their behavior under multi-hop consensus, gradient tracking, or primal–dual exchanges is unstudied. There is a concrete reason for concern: a node in a decentralized topology sends many messages per training round (one per neighbor per iteration), so the same sample's information leaks through multiple independently-defended channels, and the "local generalization" effect documented in On the (In)security of Peer-to-Peer Decentralized Machine Learning makes each of those channels individually revealing. This is a genuine threat to our proposal's necessity, and we treat it as a first-class baseline: if a local-only defense ported to the decentralized setting withstands our trajectory attack at equal utility, the added complexity of edge-pairwise noise is not justified, and we will say so.
Proposed Experiments and Decisive Checks
The attack evaluation runs on MNIST with LeNet-5 and logistic regression (EFPSN's own setting, via EFPSN and Deep Leakage from Gradients), CIFAR-10 and CIFAR-100 with LeNet-5 and ResNet-20 (the settings used by Inverting Gradients and On the (In)security of Peer-to-Peer Decentralized Machine Learning), and ImageNet with ResNet-18 to connect to the hardest reported attack setting, where High-Fidelity Gradient Inversion in Distributed Learning reports LPIPS of 0.33 and PSNR of 15.12 at batch size 8. FedNew is attacked on its LibSVM benchmarks with the tabular analogue of reconstruction (recovering feature vectors). Reconstruction quality is measured by PSNR, LPIPS, and MSE against ground-truth samples; utility is measured by test accuracy and relative optimality gap; communication by transmitted bits. The decisive checks are, first, that trajectory inversion strictly dominates per-round inversion given $K \geq n/d_b$ observed messages on the undefended protocols; second, that FedNew's composite vectors fall to the prior-guided attack at some identifiable round count, turning an untested analytical claim into a measured breaking point; third, that our defense holds test accuracy within one point of the undefended decentralized baseline while driving reconstruction below the level DP achieves at equal accuracy, across honest-component sizes $m$ from 1 to $N$ on both Poisson and power-law topologies, and under 20% message loss where we predict EFPSN's residual bias appears and ours does not. A final check on the theory side is whether measured leakage across $m$ follows the $1/(2(m-1))$ scaling of Topology-Dependent Privacy Bound for gradient-tracking messages but decays more slowly for dual trajectories — if it does, topology design alone is insufficient for operator-splitting privacy, which is the strongest form of the case this proposal is trying to make.
References
- Inverting Gradients - How easy is it to break privacy in federated learning?
- FedNew: A Communication-Efficient and Privacy-Preserving Newton-Type Method for Federated Learning
- Gradient Inversion with Generative Image Prior
- High-Fidelity Gradient Inversion in Distributed Learning
- On the (In)security of Peer-to-Peer Decentralized Machine Learning
- Topology-Dependent Privacy Bound for Decentralized Federated Learning
- Private and Accurate Decentralized Optimization via Encrypted and Structured Functional Perturbation
- Deep Leakage from Gradients
- Optimal gradient tracking for decentralized optimization
- A Distributed Operational Method for Convex Hull Pricing Based on the Alternating Direction Method of Multipliers with Dantzig–Wolfe and Benders Decomposition
- A Scalable, Distribution Network-Aware, Customer Privacy-Preserving Framework for Operation of Virtual Power Plants
- Privacy Amplification by Iteration for ADMM with (Strongly) Convex Objective Functions
- CENSOR: Defense Against Gradient Inversion via Orthogonal Subspace Bayesian Sampling
- Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks