FuSeFL: Secure Cross-Silo Federated Learning

Updated 6 July 2026

FuSeFL is a federated learning framework that decentralizes secure training by organizing clients into pairs using MPC-based techniques.
It confines the server’s role to secure aggregation, eliminating server-side training bottlenecks and ensuring both data and model confidentiality.
Empirical results show significant performance gains, including up to 25× communication speedup and competitive accuracy compared to centralized training.

Searching arXiv for papers using the term “FuSeFL” and close variants to ground the article in the relevant literature. FuSeFL is a polysemous research label rather than a single, stable technical term. In its exact expansion, it denotes “Fully Secure and Scalable Cross-Silo Federated Learning”, a federated learning scheme that decentralizes secure training across client pairs and confines the server’s role to secure aggregation (Ghinani et al., 18 Jul 2025). Closely related spellings—FuseFL, FUSE, and FUSE-Flow—name distinct frameworks in one-shot federated learning, explainable fault localization, image–event monocular depth estimation, real-time multi-camera 3D reconstruction, cognitive theory, and spectral element discretization (Tang et al., 2024, Widyasari et al., 2024, Sun et al., 25 Mar 2025, Sun, 3 Jun 2026, Jalife et al., 2021, Pan et al., 2023). The literature therefore uses the label in multiple unrelated senses, with the federated learning meaning being the exact expansion of the acronym “FuSeFL.”

1. Nomenclature and scope

The term appears in several recent arXiv papers with different expansions, objectives, and technical stacks. The exact string FuSeFL is attached to a secure cross-silo FL protocol, while nearby spellings such as FuseFL and FUSE refer to other systems.

Label	Expansion or description	Source
FuSeFL	Fully Secure and Scalable Cross-Silo Federated Learning	(Ghinani et al., 18 Jul 2025)
FuseFL	One-Shot Federated Learning through the Lens of Causality with Progressive Model Fusion	(Tang et al., 2024)
FuseFL	Explainable fault localization with LLM prompting	(Widyasari et al., 2024)
FUSE	Label-Free Image-Event Joint Monocular Depth Estimation	(Sun et al., 25 Mar 2025)
FUSE-Flow	Decoupled framework for calibration and stateless real-time multi-view point cloud fusion	(Sun, 3 Jun 2026)
fuse	“fusion of activity-related sensory stimuli and awareness”	(Jalife et al., 2021)
FUSE	Face Upwinded Spectral Element method	(Pan et al., 2023)

This multiplicity is not merely typographic. Each work introduces a different formal object: a cryptographic FL protocol, a causal OFL method, an LLM-based debugging pipeline, a self-supervised multimodal depth estimator, a stateless 3D reconstruction framework, a cognitive construct, or a numerical PDE discretization. A common misconception is therefore to treat “FuSeFL” as a single framework that migrated across domains. The papers do not support that reading.

2. FuSeFL as “Fully Secure and Scalable Cross-Silo Federated Learning”

In the exact sense of the acronym, FuSeFL is a federated learning framework that simultaneously secures client data, local updates, and the global model while eliminating the server-side training bottleneck. It is designed for cross-silo deployments in which institutions operate under strict privacy regulations (e.g., HIPAA, GDPR), have large datasets, and require confidentiality of both data and proprietary model parameters (Ghinani et al., 18 Jul 2025).

The motivating problem is twofold. First, existing secure FL methods trade security against efficiency: homomorphic encryption (HE) incurs high computational and memory costs; differential privacy (DP) degrades accuracy as the privacy budget tightens; and secure multiparty computation (MPC) protects inputs and updates without accuracy loss but sees its communication overhead grow with client count. Second, many methods leave the global model in plaintext on clients, which creates risks of model theft or misuse (e.g., extraction, reverse engineering). FuSeFL addresses both issues by keeping the model secret-shared throughout training and by moving secure training away from the server.

Its architecture is pairwise and decentralized. FuSeFL organizes clients into pairs per round, performs MPC-based training jointly within the pair over secret shares, and uses two non-colluding aggregation servers for secure aggregation only. The server role is explicitly limited: “The server’s role is strictly secure aggregation.” The global model is secret-shared so that no client or server ever holds the full global model in plaintext during training or aggregation. The paper states the security goals as data confidentiality, model confidentiality, update privacy, and integrity/correctness, under an honest-but-curious (semi-honest) threat model with no collusion assumed within a group per round (Ghinani et al., 18 Jul 2025).

The cryptographic substrate is standard MPC. For additive secret sharing, a private value satisfies

$X = \sum [x]_j.$

For secure multiplication, FuSeFL uses Beaver triples:

$E_j = X_j - a_j,\qquad F_j = Y_j - b_j$

$E = \sum E_j,\qquad F = \sum F_j$

$Z_j = j \cdot E \cdot F + F \cdot a_j + E \cdot b_j + c_j.$

For non-linearities, it uses function secret sharing (FSS) with

$f(X) = f_1(X) + f_2(X).$

Federated aggregation is performed on shares only:

$[w_{t+1}]_s = \sum_{g=1}^{G} (n_g / n) [w_t^g]_s.$

The paper also defines risk-aware dynamic grouping as a minimum-weight perfect matching

$\mathcal{G}^* = \arg\min_{\mathcal{G} \subseteq \binom{C}{2}} \sum_{(C_i,C_j)\in\mathcal{G}} \rho(C_i, C_j),$

subject to each client appearing exactly once (Ghinani et al., 18 Jul 2025).

Two execution modes are described. In FuSeFL-Serial, a pair trains a single local model share sequentially on each partner’s data. In FuSeFL-Parallel, each client trains its own local model share and the pair performs intra-pair secure aggregation before sending shares to the two aggregation servers. The serial variant prioritizes accuracy; the parallel variant prioritizes latency and throughput.

Empirically, the reported setup uses MNIST, Network-1 (3-layer MLP with ReLU), Network-2 (2 conv + 2 FC with ReLU/max-pool), and LeNet, built on AriaNN’s MPC training with two workers per group. The paper reports that FuSeFL-Serial reaches 97.20% (+0.8% over AriaNN-FL) on Network-1, matches AriaNN centralized secure training at 98.40% on Network-2, and matches AriaNN centralized at 99.20% on LeNet. At scale, with 128 clients, FuSeFL-Serial achieves up to 5.65× speedup over AriaNN-FL; FuSeFL-Parallel up to 7.83×. For communication, the paper states that for Network-1 with 10,000 clients and per-client dataset size 0.45 MB, AriaNN-FL total communication time ≈ 1,123 s, whereas FuSeFL total communication time ≈ 45.21 s, which is a 25× reduction (and up to 95% lower latency in large-scale regimes). Server memory is reduced by ≈ 50%, since each server stores (n/2) + 1 model instances rather than n local model instances + 1 global (Ghinani et al., 18 Jul 2025).

A second FL usage appears in “FuseFL: One-Shot Federated Learning through the Lens of Causality with Progressive Model Fusion” (Tang et al., 2024). This paper does not address secure MPC or model confidentiality. Instead, it addresses One-shot Federated Learning (OFL), where each client trains locally and communicates only once. Its central diagnosis is the isolation problem: under non-IID data, isolated local models in OFL can fit spurious correlations rather than invariant features.

The method is built from blockwise network decomposition and bottom-up progressive training and fusion. A network is decomposed as

$F = \Lambda \circ H^K \circ H^{K-1} \circ \cdots \circ H^1,$

and FuseFL progressively trains and fuses blocks from low to high. The feature-level fusion operator is

$fus_k(x) = \{ H_1^k(x), H_2^k(x), \dots, H_M^k(x) \},$

followed by either an Averaging adaptor

$A_{\mathrm{avg}}(x) = \mathrm{Average}(fus_k(x))$

or a 1×1 convolution adaptor

$E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 0

The paper emphasizes that this introduces “no additional communication costs” relative to OFL because summing across $E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 1 rounds equals $E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 2 per client (Tang et al., 2024).

Its empirical claims are framed as accuracy and storage trade-offs. On CIFAR-10, the paper reports FuseFL K=4: 73.79/84.58/81.15 for $E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 3, and K=2: 70.85/81.41/84.34. On SVHN, it reports K=2: 76.88/91.07/90.87, K=4: 78.08/89.63/89.34, and K=8: 80.31/88.88/89.94. The paper also states that FuseFL supports high scalability of clients, heterogeneous model training, and low memory costs, and that it is “the first attempt using causality to analyze and alleviate data heterogeneity of OFL” (Tang et al., 2024).

The relation between the two FL papers is purely nominal. FuSeFL (Ghinani et al., 18 Jul 2025) is a fully secure and scalable FL scheme designed for cross-silo settings; FuseFL (Tang et al., 2024) is a one-shot accuracy-improvement method centered on causal feature augmentation and progressive model fusion. They solve different problems and rely on different assumptions.

4. FuseFL in explainable fault localization

A third usage appears in software engineering. FuseFL in “Demystifying Faulty Code with LLM: Step-by-Step Reasoning for Explainable Fault Localization” is an LLM-assisted, explainable fault localization approach that fuses spectrum-based suspiciousness, test-case execution outcomes, and a description of intended code functionality into a structured prompt (Widyasari et al., 2024).

The method is a prompt-fusion pipeline rather than a numerical fusion algorithm. It first executes tests with coverage instrumentation, computes SBFL suspiciousness scores per line, extracts inputs and expected outputs from assert-style tests, records actual outputs or runtime errors, and then prompts ChatGPT, GPT-3.5 as of Sept-11-2023 using Zero-shot Chain-of-Thought prompting with “Let’s think step by step.” The suspiciousness formulas explicitly listed in the paper include Ochiai, Tarantula, DStar, Barinel, and OP2; for example,

$E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 4

The prompt includes the top 5 lines by suspiciousness (Widyasari et al., 2024).

The evaluation uses a curated subset of the Refactory dataset with 324 faulty Python codes, 600 faulty lines, and a new human-authored explanation set. The paper reports Top-1 localized faults of 89 for Tarantula, 117 for Ochiai, 115 for OP2, 89 for Barinel, 115 for DStar, 149 for XAI4FL, and 197 for FuSeFL. It further states “a more than 30% increase in the number of successfully localized faults at Top-1 compared to the baseline” and “FuSeFL vs XAI4FL: +32.3% at Top-1.” For explanation quality, the paper created human explanations for 600 faulty lines and reports that for 22 out of the 30 randomly sampled cases, FuseFL generated correct explanations (Widyasari et al., 2024).

This usage differs from both FL meanings. The shared term “fusion” here refers to combining debugging signals in a prompt, not to secure aggregation, pairwise MPC, or progressive feature fusion in OFL.

5. Multimodal perception and 3D reconstruction uses of FUSE

The term is also used in perception and 3D reconstruction, again with different semantics. In “FUSE: Label-Free Image-Event Joint Monocular Depth Estimation via Frequency-Decoupled Alignment and Degradation-Robust Fusion,” FUSE is a label-free (self-supervised) image–event framework for monocular depth estimation (Sun et al., 25 Mar 2025). The paper identifies two barriers: limited annotated image–event–depth datasets and frequency mismatch between modalities. Its solution combines Parameter-efficient Self-supervised Transfer (PST) and Frequency-Decoupled Fusion (FreDFuse).

PST is a two-stage, lightweight transfer that aligns an event encoder’s latent space with an image depth foundation model such as DepthAnything V2. Stage I updates only LoRA and PatchEmbed in the event encoder, reported as about 1.8% of the encoder’s parameters. Stage II freezes the encoders and trains only FreDFuse on randomly degraded image–event inputs. The alignment loss is

$E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 5

with a cosine-gated latent term and the similarity band $E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 6. Events are voxelized into $E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 7 with the paper uses $E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 8 (Sun et al., 25 Mar 2025).

FreDFuse explicitly separates high-frequency components (edges, motion dynamics) from low-frequency components (global structure, smooth surfaces) using Gaussian–Laplacian pyramids, then performs modality-guided cross-attention in which the event branch dominates high-frequency fusion, and the image branch dominates low-frequency fusion. The paper reports state-of-the-art performance with 14% and 24.9% improvements in Abs.Rel on MVSEC and DENSE datasets, FreDFuse (vs. vanilla cross-attention) cuts Abs.Rel by ~8%, and PST reduces trainable parameters by >80% while improving average performance by ~19.7% across metrics (Sun et al., 25 Mar 2025).

A separate but related usage appears in “FUSE-Flow: A Decoupled Framework for Calibration and Stateless Real-Time Multi-View Point Cloud Fusion” (Sun, 3 Jun 2026). FUSE-Flow comprises GMAC (Geometry-aligned Multi-view extrinsic calibration) and FUSE (Reliability-guided multi-view point cloud fusion). The framework is explicitly decoupled and stateless: no TSDF volume, SLAM pose graph, or bundle adjustment is maintained. GMAC performs metric lifting of transformer-predicted poses using Depth-guided Reprojection Consistency (RC) and Cycle Consistency (CC); FUSE uses Measurement Confidence Modeling (MCM), Cross-View Geometric Consistency (CVGC), and Adaptive Spatial Hashing (ASH).

The calibration objective is written as

$E_j = X_j - a_j,\qquad F_j = Y_j - b_j$ 9

with the closed-form scale estimator

$E = \sum E_j,\qquad F = \sum F_j$ 0

The fusion weights use the harmonic mean

$E = \sum E_j,\qquad F = \sum F_j$ 1

The paper reports that FUSE achieves 26.3–43.1 FPS across 2–8 cameras at 640×480 on an RTX 4090, uses 1.9–3.1 GB GPU, and that R3D3 runs at 5.8–11.2 FPS with 6.4–10.1 GB for comparable settings. For calibration, VGGT 1.04° → 0.63°, 4.61 mm → 3.07 mm on ScanNet is reported as the best absolute among the listed backbones (Sun, 3 Jun 2026).

These FUSE systems are unrelated to FuSeFL in federated learning, but they illustrate how the same or adjacent acronymic forms recur in multimodal fusion problems.

6. Cognitive and numerical-scientific uses

The word fuse in “From Flow to Fuse: A Cognitive Perspective” is not an acronym. It is introduced as a new construct defined as the “fusion of activity-related sensory stimuli and awareness” (Jalife et al., 2021). The paper argues that the prevalent notion of flow is conceptually vague, especially around the phrase “merging of action and awareness.” Fuse narrows the phenomenology to episodes in which consciousness is filled exclusively with activity-related sensory stimuli, while planning, model-building, mental simulation, conscious comparison of alternatives do not enter awareness.

The paper’s preliminary model is built around working memory (WM), attention, saliency, automaticity, mind wandering/DMN, and cognitive load. It specifies a requirement—Effortless attention focused on activity-related sensory stimuli—and three corequisites: Effortless avoidance of activity-related thoughts, Effortless avoidance of activity-unrelated thoughts, and Effortless avoidance of activity-unrelated sensory stimuli. It then derives design heuristics such as Minimize time between actions, Minimize relatability, Leverage sound design, Help individuals meet their OCLs, Optimize complexity, not difficulty, and Facilitate learning with predictable systems (Jalife et al., 2021).

In numerical analysis, FUSE names the Face Upwinded Spectral Element method (Pan et al., 2023). The paper introduces a high-order, linearly stable discretization on unstructured quadrilateral meshes that keeps the solution field continuous (CG-like) while incorporating upwind stabilization only on face nodes. The key node placement is the Gauss–Legendre quadrature nodes plus endpoints (“GL+E”). The paper states that equidistant or Gauss–Lobatto nodes lead to instability for $E = \sum E_j,\qquad F = \sum F_j$ 2, whereas GL+E nodes yield a stable upwind operator (verified numerically up to $E = \sum E_j,\qquad F = \sum F_j$ 3). For the semi-discrete 1D advection system, stability is characterized by $E = \sum E_j,\qquad F = \sum F_j$ 4 for all eigenvalues of the derivative operator. The method is demonstrated on convection-dominated flows, Poisson’s equation, and the incompressible Navier–Stokes equations, with empirical convergence between orders $E = \sum E_j,\qquad F = \sum F_j$ 5 and $E = \sum E_j,\qquad F = \sum F_j$ 6 (Pan et al., 2023).

The paper explicitly notes a terminological caveat: “The name ‘FuSeFL’ does not appear in the paper; we treat ‘FuSeFL’ as referring to the same Face Upwinded Spectral Element approach described therein.” That statement is important because it shows that, in some discussions, the label is being extended beyond the authors’ own nomenclature.

7. Sources of ambiguity and comparative interpretation

Across the cited works, the label family FuSeFL / FuseFL / FUSE / FUSE-Flow / fuse spans at least four distinct patterns. First, there are federated learning uses: secure cross-silo training with secret sharing and MPC (Ghinani et al., 18 Jul 2025), and one-shot FL with progressive blockwise model fusion (Tang et al., 2024). Second, there are signal-fusion and multimodal perception uses: prompt-level fusion for fault localization (Widyasari et al., 2024), frequency-decoupled image–event fusion (Sun et al., 25 Mar 2025), and reliability-guided multi-view point cloud fusion (Sun, 3 Jun 2026). Third, there is the cognitive use of fuse as a phenomenological construct (Jalife et al., 2021). Fourth, there is the numerical PDE use of FUSE as a face-upwinded discretization (Pan et al., 2023).

This suggests that acronym reuse is often driven by the mnemonic appeal of fuse/fusion, but the technical content is not shared across papers. That generalization is strongest for (Tang et al., 2024, Widyasari et al., 2024, Sun et al., 25 Mar 2025), and (Sun, 3 Jun 2026), all of which explicitly describe some form of fusion. A plausible implication is that the proximity of these labels increases the risk of bibliographic or conceptual confusion, especially because the exact string FuSeFL has a specific meaning in (Ghinani et al., 18 Jul 2025) but is also used informally in summaries of unrelated work.

For precision, the term should therefore be disambiguated by domain. In federated learning, FuSeFL most precisely denotes Fully Secure and Scalable Cross-Silo Federated Learning (Ghinani et al., 18 Jul 2025). In one-shot FL, FuseFL denotes progressive model fusion for causal feature augmentation (Tang et al., 2024). In software engineering, FuseFL denotes an LLM-assisted explainable fault localization method (Widyasari et al., 2024). In perception and reconstruction, FUSE and FUSE-Flow denote multimodal fusion systems (Sun et al., 25 Mar 2025, Sun, 3 Jun 2026). In cognitive theory, fuse denotes a redefinition of a class of sensory-bound experiences (Jalife et al., 2021). In numerical analysis, FUSE denotes the Face Upwinded Spectral Element method (Pan et al., 2023).