Federated Flow Matching (FFM)

• Federated Flow Matching is a privacy-preserving generative modeling approach that trains flow matching models across decentralized datasets without centralizing raw data.
• It features distinct variants—FFM-vanilla, FFM-LOT, and FFM-GOT—which balance tradeoffs between simplicity, local optimal transport, and global geodesic performance.
• FFM underpins robust federated learning frameworks like FedFG, enabling feature synthesis for adversary detection while maintaining strict data privacy.

Federated Flow Matching (FFM) is a class of privacy-preserving generative modeling and learning techniques that enable the training of flow matching models across decentralized datasets under federated learning (FL) constraints. These methods are designed to reconcile strong privacy—ensuring that raw data remains strictly local and never transmitted—with high-quality generative performance and, in extensions, robustness to adversarial attack. FFM encompasses both foundational algorithms for federated generative modeling (Wang et al., 25 Sep 2025) and robust, privacy-preserving FL frameworks that leverage flow-based feature synthesis rather than direct parameter or feature sharing (Wang et al., 30 Mar 2026).

1. Theoretical Motivation and Problem Setting

Federated Flow Matching is motivated by the need to perform generative modeling or collaborative learning when data is distributed across clients (e.g., mobile devices, hospitals) and cannot be centralized due to privacy, ownership, or regulatory concerns. Classical flow-matching generative models define a parameterized vector field $v_t$ that transports a simple distribution $q_0$ (e.g., Gaussian noise) to a complex target $q_1$ (e.g., natural images), minimizing the regression loss

$$\mathcal L_{\mathrm{CFM}}(\theta) = \mathbb E_{t\sim\mathcal U[0,1],\,(x_0,x_1)\sim\pi} \Big\|v_t^\theta((1-t)x_0 + t x_1)\;-\;(x_1 - x_0)\Big\|^2,$$

where $\pi$ is a coupling between $q_0$ and $q_1$. In federated settings, each client holds a private local target distribution $q_1^i$, with the global target being the mixture $q_{1,\lambda} = \sum_{i=1}^n\lambda_i q_1^i$. The central challenge is to build effective couplings and flow-matching objectives to learn generative models aligned with the global mixture, without ever centralizing data (Wang et al., 25 Sep 2025).

For privacy-preserving federated learning (beyond pure generative modeling), federated flow-matching techniques provide a mechanism for synthesizing feature representations that enable robust aggregation and adversary detection, despite never revealing private feature extractors or raw data (Wang et al., 30 Mar 2026).

2. Federated Flow Matching Algorithms and Mathematical Formulations

Federated Flow Matching is structured around three principal algorithms, each of which resolves the privacy/coupling tradeoff differently:

Variant	Coupling Construction	Principal Features
FFM-vanilla	Independent product	Simple, stable; curved flows
FFM-LOT	Local OT per client	Local straightness; lacks global
FFM-GOT	Global OT (semi-dual)	Geodesic flows; maximal fidelity

2.1. FFM-vanilla: Uses the independent product coupling $\pi_{\mathrm{vanilla}} = q_0 \otimes q_{1,\lambda}$, yielding a loss

$q_0$0

computed and communicated locally, preserving privacy but resulting in curved flows and slower inference (Wang et al., 25 Sep 2025).

2.2. FFM-LOT: Each client computes its own local optimal transport plan $q_0$1 between $q_0$2 and $q_0$3, and the server aggregates $q_0$4. This variant achieves straighter flows locally but diverges from the true global OT coupling under heterogeneity, resulting in sub-optimal aggregate flows (Wang et al., 25 Sep 2025).

2.3. FFM-GOT: Approximates the global optimal transport plan via the Kantorovich semi-dual. Each client participates in optimizing a shared potential $q_0$5 using local data, exchanging only gradients via federated averaging. The coupling is indirectly constructed using $q_0$6-transforms and candidate pools: $q_0$7 Sampling pairs is implemented through minimization over a pool of source candidates for each target $q_0$8 (Wang et al., 25 Sep 2025). This variant yields nearly straight displacement flows (true geodesics) and best matches the performance of centralized OT-based training, particularly at low inference step counts.

3. Privacy-Preserving and Robust Federated Learning via Flow Matching

Frameworks such as FedFG (Wang et al., 30 Mar 2026) build on federated flow matching principles to provide robust and privacy-preserving supervised federated learning. In FedFG, the client network is decomposed as follows:

• Private extractor $q_0$9, retained locally.
• Public classifier $q_1$0, parameters shared.
• Conditional flow-matching generator $q_1$1, shared. Internally, $q_1$2 is parameterized by a vector field $q_1$3 and used to generate synthetic feature samples via integrating an ODE:

$q_1$4

The generator learns to transport Gaussian noise to the distribution of private features, conditioned on labels.

Client-side training alternates between:

1. Freezing $q_1$5, optimizing $q_1$6 via SGD on the classification loss

$q_1$7

2. Optimizing $q_1$8 via SGD on

$q_1$9

where $$\mathcal L_{\mathrm{CFM}}(\theta) = \mathbb E_{t\sim\mathcal U[0,1],\,(x_0,x_1)\sim\pi} \Big\|v_t^\theta((1-t)x_0 + t x_1)\;-\;(x_1 - x_0)\Big\|^2,$$0 is the usual flow-matching loss regressing the local vector field to $$\mathcal L_{\mathrm{CFM}}(\theta) = \mathbb E_{t\sim\mathcal U[0,1],\,(x_0,x_1)\sim\pi} \Big\|v_t^\theta((1-t)x_0 + t x_1)\;-\;(x_1 - x_0)\Big\|^2,$$1 along interpolated paths.

Only $$\mathcal L_{\mathrm{CFM}}(\theta) = \mathbb E_{t\sim\mathcal U[0,1],\,(x_0,x_1)\sim\pi} \Big\|v_t^\theta((1-t)x_0 + t x_1)\;-\;(x_1 - x_0)\Big\|^2,$$2 and $$\mathcal L_{\mathrm{CFM}}(\theta) = \mathbb E_{t\sim\mathcal U[0,1],\,(x_0,x_1)\sim\pi} \Big\|v_t^\theta((1-t)x_0 + t x_1)\;-\;(x_1 - x_0)\Big\|^2,$$3 are transmitted to the server; $$\mathcal L_{\mathrm{CFM}}(\theta) = \mathbb E_{t\sim\mathcal U[0,1],\,(x_0,x_1)\sim\pi} \Big\|v_t^\theta((1-t)x_0 + t x_1)\;-\;(x_1 - x_0)\Big\|^2,$$4 remains private