Domain Feature Corrector (DFC) in Federated Learning

• DFC is a residual corrector in the F²DC framework that calibrates domain-related features using direct classification supervision.
• It leverages a two-layer CNN to refine, rather than discard, mixed semantic information affected by domain bias.
• Empirical evaluations, including ablation studies and t-SNE visualizations, show that DFC enhances cross-domain consistency and representation quality.

Searching arXiv for the cited DFC-related papers to ground the article and disambiguate the term. Domain Feature Corrector (DFC) is a feature-level module introduced in the federated learning framework $F^2$DC for domain-skewed learning across clients from different domains (Wang et al., 15 Mar 2026). Its function is to calibrate domain-related features rather than remove them: after a preceding decoupling stage separates local representations into domain-robust and domain-related parts, DFC takes the domain-related part and recovers class-relevant clues that remain entangled with domain bias. In that formulation, DFC is a residual corrector trained by direct classification supervision so that the corrected domain-related representation complements the domain-robust representation and improves cross-domain consistency (Wang et al., 15 Mar 2026).

1. Problem setting and motivation

The DFC formulation arises in federated learning under domain skew, where clients originate from different domains but have roughly similar label distributions. The paper characterizes this setting by a mismatch in conditional feature distributions across clients,

$$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$

Under this condition, local models tend to overfit domain-specific context, and the resulting local representations collapse into a narrow low-dimensional subspace, which degrades consensus under server aggregation and harms generalization across domains (Wang et al., 15 Mar 2026).

DFC is motivated by a specific critique of elimination-based domain generalization. Many prior methods attempt to remove domain-specific features. The $F^2$DC paper argues that this can be overly aggressive because domain-related features often contain not only nuisance bias but also useful class-relevant evidence, such as object contours or parts (Wang et al., 15 Mar 2026). DFC is therefore introduced not as a denoising filter that discards domain-related information, but as a calibration mechanism that repairs the useful portion of that information.

This implies a precise conceptual distinction. Domain-robust features are treated as the primary cross-domain semantic substrate, whereas domain-related features are treated as mixed carriers of both bias and residual semantic evidence. DFC exists because the latter are not assumed to be useless.

2. Position within the $F^2$DC framework

DFC is the second stage of the local feature remedy in $F^2$DC. For a sample $(x_i,y_i)$ on client $k$, the backbone produces a feature map $f_i \in \mathbb{R}^{C\times H\times W}$. The preceding Domain Feature Decoupler (DFD) applies an attribution module $\mathcal{A}_D$ to compute a robustness score map,

$$\mathcal{S}_i = \mathcal{A}_D(f_i) \in \mathbb{R}^{C\times H\times W},$$

and then constructs a pseudo-binary mask $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$0 using a Gumbel-concrete relaxation so that the selection remains differentiable (Wang et al., 15 Mar 2026).

The feature split is

$$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$1

Here $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$2 is the domain-robust feature and $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$3 is the domain-related feature. DFC takes $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$4 as input (Wang et al., 15 Mar 2026).

Within the full pipeline, local feature decoupling is followed by local feature correction and then by domain-aware aggregation. The server performs aggregation using weights that depend on both sample size and a domain discrepancy score, but the decoupler $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$5, corrector $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$6, and classifier $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$7 remain local and are not aggregated (Wang et al., 15 Mar 2026). That locality is significant because DFC is designed as a client-specific repair operator rather than a globally shared calibration head.

3. Architectural form and correction mechanism

DFC is implemented as a Domain Feature Corrector $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$8 that transforms the domain-related feature $$\mathbb{P}_{k_1}(x \mid y) \neq \mathbb{P}_{k_2}(x \mid y), \quad \text{s.t.} \quad \mathbb{P}_{k_1}(y)=\mathbb{P}_{k_2}(y).$$9 into a corrected feature. Architecturally, it uses the same two-layer CNN as DFD, with batch normalization and ReLU (Wang et al., 15 Mar 2026).

The defining mechanism is residual correction:

$F^2$0

This residual design means that the corrector does not overwrite the original domain-related signal; it refines only the portion identified by $F^2$1 as domain-related (Wang et al., 15 Mar 2026). The paper presents this as a flexible and stable design.

The authors describe DFC as calibrating domain-related features “by explicitly linking correct semantic signals” (Wang et al., 15 Mar 2026). Operationally, this means that the domain-related stream is not trusted as-is. Instead, the corrector learns to inject back class-aligned signals into that stream. The target is not domain invariance in the strict sense; it is semantic recovery inside a feature branch that contains both domain artifacts and mixed class information.

The corrected feature is explicitly complementary to the domain-robust feature. The final representation is formed as

$F^2$2

This additive composition captures the central design philosophy of $F^2$3DC: separate first, then correct what was separated out (Wang et al., 15 Mar 2026).

4. Objective function and semantic supervision

DFC is supervised directly by classification rather than by a contrastive or reconstruction objective. The DFC loss is

$F^2$4

where $F^2$5 is the embedding of the corrected feature after the pooling and flattening layer $F^2$6, $F^2$7 is a single-layer MLP classifier, and $F^2$8 is the softmax (Wang et al., 15 Mar 2026). The paper emphasizes that this is what “explicitly linking discriminative signals” means in practice: the corrected feature is required to predict the true class.

DFC is part of a larger local objective,

$F^2$9

The standard cross-entropy term is

$F^2$0

and the DFD term creates the decomposition on which DFC relies (Wang et al., 15 Mar 2026).

This training arrangement formalizes a division of labor. DFD enforces separability between robust and domain-related features while retaining discriminative ability; DFC then takes the less reliable branch and makes it useful instead of discarding it. A common misconception is therefore that DFC serves the same role as the robust branch. The paper is explicit that it does not. Domain-robust features $F^2$1 are meant to capture cross-domain class evidence that should generalize well, whereas DFC converts $F^2$2 into corrected features $F^2$3 that recover useful clues mixed with bias (Wang et al., 15 Mar 2026).

5. Empirical characterization

The empirical evidence reported for DFC is primarily ablation-based. In Table 6, removing DFC while keeping other components lowers performance relative to the full system; the complete model is best, indicating that DFC adds value beyond DFD alone (Wang et al., 15 Mar 2026). Table 7 further compares the robust feature $F^2$4, the domain-related feature $F^2$5, a corrected feature $F^2$6, and the final combined feature $F^2$7. The reported pattern is consistent across settings: $F^2$8 is much better than $F^2$9, $F^2$0 is consistently better than $F^2$1, and $F^2$2 is best overall (Wang et al., 15 Mar 2026).

The qualitative evidence is aligned with that ablation pattern. Grad-CAM visualizations show that $F^2$3DC attends more consistently to semantic object regions across domains, and the corrected features recover missing object parts that elimination-based methods miss. The paper gives a PACS example in which DFD may place the giraffe’s upper body in the robust branch and leave the rest in the domain-related branch, while DFC recovers additional cues such as the waist (Wang et al., 15 Mar 2026).

The authors also report that final representations cluster better in t-SNE plots, with tighter within-class grouping and larger between-class separation (Wang et al., 15 Mar 2026). Their interpretation is that DFC improves representation consistency by bringing domain-related cues back into alignment with class semantics. This suggests that DFC contributes both within-domain and across-domain: within a client, it enables the model to use more of the available evidence; across clients, it reduces the extent to which domain-related features remain tied to a particular domain style.

A further experimental observation is that too aggressive separation harms performance, which indirectly supports DFC’s role. If DFD tries too hard to push all semantics into the robust branch, performance degrades; DFC helps because it exploits the residual class information left behind in the domain-related branch (Wang et al., 15 Mar 2026).

6. Conceptual significance and boundaries of the term

Within federated learning, DFC represents a correction-based alternative to elimination-only treatments of domain bias. Its conceptual claim is not that domain-related information should be preserved indiscriminately, but that it should be calibrated when it contains mixed semantic content. That stance distinguishes $F^2$4DC from comparison baselines such as FDSE, which attempt to discard domain bias and may also discard useful class evidence (Wang et al., 15 Mar 2026).

At the same time, the acronym “DFC” is overloaded in the arXiv literature. In nonlinear dynamics, DFC denotes delayed feedback control rather than Domain Feature Corrector (Dmitrishin et al., 2014). In remote-sensing change detection, DFC denotes detail feature compensation, a decoder-side module for compensating detail loss and contamination during FPN upsampling (Xiao et al., 2024). In point cloud registration, DFC denotes deep feature consistency, a registration framework built around a feature consistency matrix and weighted Procrustes estimation (Xu et al., 2021). These usages are unrelated in method and objective, and the federated-learning DFC should therefore be identified by its full name when ambiguity is possible.

A plausible implication is that the federated-learning DFC belongs to a broader family of corrective strategies for distributional variation. In a different setting, predictor-corrector methods for stochastic optimization under gradual distribution shift show that explicitly modeling the direction of drift can outperform purely reactive adaptation (Maity et al., 2022). That work is not about Domain Feature Corrector, but it provides a mathematically adjacent perspective on why a corrective stage may be preferable to simple re-fitting under evolving domains.

In its original sense, however, Domain Feature Corrector is specifically the residual CNN module in $F^2$5DC that calibrates domain-related features by direct class supervision so that they complement domain-robust features under domain-skewed federated learning (Wang et al., 15 Mar 2026).