Discriminative Expert Array (DEA)

• Discriminative Expert Array is a structured ensemble dividing tasks across manifold branches and granularity levels to replace a single feature space.
• It enhances synthetic image detection by boosting cross-generator generalization and robustness against post-processing methods.
• DEA inspires related architectures in aerial object and biomedical data detection, leveraging expert-like selection and distinctive element analysis.

Discriminative Expert Array (DEA) denotes, in its most explicit usage, a structured expert ensemble that replaces a single universal discriminative feature space with many small, specialized discriminative subspaces for synthetic image detection. In TrueMoE, DEA is organized along the two axes of manifold structure and perceptual granularity, so that each expert is specialized to a unique combination of manifold branch and granularity level (Zhang et al., 19 Sep 2025). Closely related but distinct usages appear in aerial object detection, where Dynamic Enhancement Anchor is described as behaving like a discriminative expert array through discriminator-guided interaction between anchor-based and anchor-free sample generators (Liang et al., 2021), and in biomedical data exploration, where Distinctive Element Analysis is conceptually treated as an array of discriminative experts built from distinctive elements of high-dimensional datasets (Islam et al., 2024).

1. Terminology and scope

The initials “DEA” are not tied to a single expansion across the cited literature. Instead, they designate different technical constructs that share a family resemblance: multiple specialized units, a selection or ranking mechanism, and a final discriminative output.

Usage	Meaning	Domain
Discriminative Expert Array	Structured expert ensemble over manifold structure and perceptual granularity	Synthetic image detection
Dynamic Enhancement Anchor	Anchor-based and anchor-free interaction behaving like a discriminative expert array	Aerial object detection
Distinctive Element Analysis	Unsupervised extraction and condensation of distinctive data elements, conceptually viewed as an expert array	Biomedical datasets

In TrueMoE, the term is formal and architectural: DEA is the central construct of a dual-routing mixture-of-discriminative-experts framework, arranged as a two-dimensional grid of experts $E_{i,j}$ (Zhang et al., 19 Sep 2025). In the aerial detection work, the formal acronym remains “Dynamic Enhancement Anchor,” but the design is explicitly interpreted as a discriminative expert array because one branch is effective for small, hard-to-anchor objects, another is effective for large or extreme-aspect-ratio objects, and a discriminator selects between them per object (Liang et al., 2021). In the biomedical work, the formal acronym is “distinctive element analysis,” yet the details explicitly note that each centroid and its associated distance pattern can be viewed as an “expert,” with the final representation acting as a condensed array of discriminative experts (Islam et al., 2024).

2. Canonical formulation in TrueMoE

In TrueMoE, DEA is defined as a structured expert ensemble “organized along two synergistic axes: manifold structure and perceptual granularity,” with experts arranged in a two-dimensional grid (Zhang et al., 19 Sep 2025). Formally,

$$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$

Here, $i$ indexes manifold branches and $j$ indexes granularity levels.

Each forgery discriminative expert $E_{i,j}$ contains three components: an encoder-decoder pair $(\mathcal{E}_i,\mathcal{D}_i)$, a Hierarchical Granularity Encoder $\mathcal{H}$, and a Discriminative Head $\mathcal{P}$. The encoder-decoder pair produces a manifold-specific latent code and reconstruction,

$$z_i = \mathcal{E}_i(x), \quad x'_i = \mathcal{D}_i(z_i),$$

while the granularity branch compares the original image and reconstruction at a selected HGE depth. The Granularity-Aware Discrimination Feature is

$$\text{GDF}_j(x) = \text{Agg}\left(\mathcal{H}(x)_j,\mathcal{H}(x'_i)_j\right),$$

with $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$0. The residual is described as highlighting subtle inconsistencies indicative of forgery. The expert output is then

$$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$1

The two organizational axes are instantiated concretely. Along the manifold axis, each branch corresponds to a separate autoencoder pretrained on a particular generative model. The implementation uses three manifold branches initialized from autoencoders trained on Kandinsky 2.1, Stable Diffusion 1.1, and Stable Diffusion 2.1, so $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$2. Along the granularity axis, the HGE uses multiple layers of a pre-trained VGG-16; shallow layers are sensitive to low-level textures, while deeper layers capture higher-level inconsistencies. The paper uses six granularity levels, so $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$3. The default DEA therefore contains $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$4 experts (Zhang et al., 19 Sep 2025).

A notable systems choice is that only the discriminative head is trainable, while the autoencoders and HGE remain frozen. This yields lightweight experts whose specialization is induced by the manifold branch, the selected perceptual depth, and the residual feature definition rather than by end-to-end retraining of heavy backbones.

3. Routing, specialization, and optimization

DEA in TrueMoE is activated by a dual-routing mechanism composed of a Granularity Routing Module (GRM) and a Hybrid Manifold Routing Module (HMRM) (Zhang et al., 19 Sep 2025). The GRM performs sparse routing in granularity. It uses a Granularity-Aware Artifact Extractor trained with a CLIP-style contrastive objective, then clusters granularity-aware embeddings via K-means into six centers. At inference, GRM applies sparse top-$$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$5 routing with $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$6, selecting a single granularity level $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$7. This activates only the expert subgroup $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$8.

The HMRM then performs dense routing across manifold branches within the selected granularity group. It relies on a Multi-Domain Latent Representation Extractor with RGB, SRM, and DFT branches. After reconstruction pretraining, the decoders are discarded, the encoders are frozen, and their outputs are concatenated to form a manifold-sensitive representation. A manifold mapping network and routing gate convert this representation into soft weights $$\text{DEA} = \{ E_{i,j} \mid i = 1,\dots,M,\; j = 1,\dots,G \}.$$9 over the manifold branches. The aggregated decision can be written as

$i$0

Training is correspondingly structured. Expert heads are first jointly pretrained on the full ProGAN training set with a binary cross-entropy detection loss. Granularity specialization then proceeds by K-means partitioning of images into six clusters $i$1, after which the corresponding expert subsets are fine-tuned on their assigned clusters. Throughout this process, the autoencoders, HGE, GAE, and MLRE encoders remain frozen. Routing is trained with an uncertainty-aware loss and a load-balancing regularizer, and the overall objective is

$i$2

with $i$3 and $i$4 (Zhang et al., 19 Sep 2025).

This training regime is significant because it separates specialization into three mechanisms: architectural bias from distinct autoencoders and HGE layers, contrastive pretraining for granularity, and routing regularization for stable and balanced expert use.

4. Empirical behavior in synthetic image detection

TrueMoE reports that replacing a single discriminative space with DEA improves both cross-generator generalization and robustness to post-processing (Zhang et al., 19 Sep 2025). Against strong baselines such as CNNSpot, UniFD, FatFormer, and CO-SPY, the reported aggregate performance is $i$5 mean AP and $i$6 mean Accuracy, compared with $i$7 and $i$8 for the best baseline, CO-SPY. Selected AP values include 97.40 on ADM, 98.36 on SD v1.4, 98.40 on SD v1.5, and 93.28 on Midjourney.

Robustness results are also explicit. Under blurring, cropping, JPEG compression, and additive noise, TrueMoE achieves 86.3 / 91.1, 85.6 / 89.9, 87.2 / 90.3, and 85.0 / 89.2 in Acc/AP, respectively. The corresponding CO-SPY values are 83.7 / 88.3, 83.3 / 87.7, 84.2 / 87.5, and 82.1 / 86.5. The paper attributes this behavior to the large and diverse detection space constructed by DEA, together with routing features that suppress semantic content and emphasize forgery-relevant structure.

Ablation studies isolate the contribution of the array structure. Residual aggregation in the GDF is superior to concatenation, original-only, or reconstruction-only features, with 90.7 / 97.5 versus 88.6 / 96.3, 83.1 / 89.1, and 72.6 / 78.4 in Acc/AP. Increasing the number of granularity levels from 1 to 6 raises performance from 81.3 / 87.3 to 90.7 / 97.5. Increasing manifold branches from 1 to 3 raises performance from 85.3 / 89.0 to 90.7 / 97.5, while 4 branches gives 91.0 / 96.9, indicating a small AP drop relative to the default 3-branch configuration. Further ablations show that the granularity-aware contrastive feature extractor outperforms vanilla CLIP, ResNet-50, and fine-tuned CLIP features, and that multi-domain MLRE inputs outperform RGB-only, SRM-only, or DFT-only variants (Zhang et al., 19 Sep 2025).

These findings establish DEA, in the TrueMoE sense, as a domain-specific mixture-of-experts design in which the expert grid, the routing scheme, and the residual feature definition are all empirically material.

5. DEA as Dynamic Enhancement Anchor in aerial object detection

In aerial object detection, “DEA” formally means Dynamic Enhancement Anchor rather than Discriminative Expert Array, but the system is explicitly described as behaving like a discriminative expert array (Liang et al., 2021). DEA-Net integrates a conventional anchor-based RPN, an FCOS-style anchor-free head, and a sample discriminator that compares anchor-based proposals and anchor-free regressed boxes against each ground-truth object.

The detector uses backbone plus FPN, an anchor-based module with 3 horizontal anchors per level and aspect ratios $i$9, an anchor-free module that predicts

$j$0

and a discriminator with positive threshold $j$1 and negative threshold $j$2. For each ground-truth box $j$3, the discriminator compares IoUs from anchor-free boxes $j$4 and anchor-based proposals $j$5. If the anchor-free candidate has higher IoU and exceeds the positive threshold, it is added to the enhancement set $j$6; if the anchor-based candidate is better and exceeds the positive threshold, it is added to $j$7; low-IoU anchors become negatives in $j$8. Enhanced samples are then merged into the positives.

The paper interprets this as an expert-like sample generator. The anchor-free unit is good at small, hard-to-anchor objects, while the anchor-based unit is good at large or extreme-aspect-ratio objects. The discriminator selects which branch supplies the training sample for each object. During inference, however, the system uses a conservative anchor-based inference scheme: the anchor-free branch is frozen and only the anchor-based RPN plus ROI head are used. This is a training-time expert array rather than an inference-time ensemble.

Reported gains are concrete. On DOTA oriented detection, DEA-Net integrated with RoI-Transformer and ResNet-101 reaches $j$9 mAP, exceeding R$E_{i,j}$0Det-DCL by $E_{i,j}$1 despite using ResNet-101 rather than ResNet-152. On DOTA horizontal detection, it reaches $E_{i,j}$2 mAP, surpassing SCRDet by $E_{i,j}$3 with the same backbone. ReDet + DEA reaches $E_{i,j}$4 mAP. On HRSC2016, DEA-Net reaches $E_{i,j}$5 mAP, surpassing the previous best model by $E_{i,j}$6 using only 3 horizontal anchors. The paper also reports that adding more small anchors is less favorable than adding DEA: $E_{i,j}$7 versus $E_{i,j}$8 mAP, with higher GFLOPs and lower FPS for the extra-anchor alternative (Liang et al., 2021).

6. DEA as Distinctive Element Analysis in biomedical data

In biomedical data exploration, DEA denotes Distinctive Element Analysis, an unsupervised deep learning framework that extracts the distinctive data elements of high-dimensional biomedical and clinical datasets (Islam et al., 2024). The details explicitly note that the method can be understood conceptually as a discriminative expert array, where each centroid and its associated distance pattern behaves like an expert focused on a specific data element.

The framework is a kernel-driven triple-optimization network. The first optimization computes many cluster centroids $E_{i,j}$9 in the original feature space and constructs a distance matrix $(\mathcal{E}_i,\mathcal{D}_i)$0 from correlation distances between centroids and data points. The second optimization builds a kernel similarity matrix

$(\mathcal{E}_i,\mathcal{D}_i)$1

optimizes over $(\mathcal{E}_i,\mathcal{D}_i)$2, ranks centroids by importance, and keeps only the top fraction, typically $(\mathcal{E}_i,\mathcal{D}_i)$3. The third optimization trains an encoder-decoder on the filtered representation, with encoder

$(\mathcal{E}_i,\mathcal{D}_i)$4

and reconstruction loss augmented by L2 and sparsity regularization. The resulting latent variables are the DEA components.

The implementation details are specific: maximum epochs 20,000, encoder activation logistic sigmoid, decoder activation saturated linear, $(\mathcal{E}_i,\mathcal{D}_i)$5, sparsity proportion $(\mathcal{E}_i,\mathcal{D}_i)$6, sparsity coefficient $(\mathcal{E}_i,\mathcal{D}_i)$7, $(\mathcal{E}_i,\mathcal{D}_i)$8 explored from 0.5 to 1.5 in steps of 0.1, and $(\mathcal{E}_i,\mathcal{D}_i)$9 neighbors in the kernel stage. The method is used as an unsupervised preprocessor whose components are then classified by MATLAB’s multiclass ECOC SVM.

Applications include diabetic retinopathy images, ChestX-ray14, and single-cell RNA-seq. The abstract reports that DEA improves accuracy by up to $\mathcal{H}$0 compared to traditional techniques. More specific results include more than $\mathcal{H}$1 accuracy improvement over PCA and NNMF on diabetic retinopathy with 8 components and only 30% of the data for training, and at least $\mathcal{H}$2 improvement over PCA and NNMF on chest X-ray classification with 32 components. In scRNA-seq, DEA top-30 genes render all 15 retinal bipolar cell types separable in t-SNE visualizations, whereas PCA and NNMF top-gene selections do not separate classes clearly. The framework also permits user-guided manipulation of intermediate results, including manual removal of background or noise components, which the paper reports can further improve classification accuracy (Islam et al., 2024).

7. Cross-domain interpretation, misconceptions, and limitations

The cited literature suggests a recurring design pattern behind the phrase “Discriminative Expert Array,” even where it is not the formal acronym. In TrueMoE, the pattern is explicit: a two-dimensional grid of experts, sparse granularity routing, dense manifold routing, and weighted expert collaboration (Zhang et al., 19 Sep 2025). In DEA-Net, the same logic appears as discriminator-guided per-object selection between two specialized sample generators, one anchor-based and one anchor-free (Liang et al., 2021). In Distinctive Element Analysis, the pattern appears as a large set of centroid-based elements that are filtered, ranked, and condensed into a smaller discriminative latent representation (Islam et al., 2024).

A common misconception would be to treat DEA as a single standardized term across machine learning. The literature block instead shows three technically distinct uses. The more precise interpretation is that “Discriminative Expert Array” is a formal architecture in synthetic image detection, but also a useful conceptual description for systems that distribute discrimination across multiple specialized subspaces, branches, or elements.

The limitations are likewise domain-specific. TrueMoE explicitly notes the need to reduce training costs, enhance scalability, and better handle extremely subtle forgeries (Zhang et al., 19 Sep 2025). DEA-Net notes that for objects with extreme aspect ratios, both anchor-based and anchor-free branches can fail, so the discriminator cannot recover a good sample if neither branch yields good IoU; it also notes that improvements are more pronounced on horizontal than oriented detection and that the method is designed and tuned for aerial images (Liang et al., 2021). Distinctive Element Analysis assumes that datasets contain distinctive elements between groups and is therefore not suited to single-class exploration; it is also computationally more expensive than PCA or NNMF because of clustering, kernel optimization, and encoder-decoder training (Islam et al., 2024).

Taken together, these formulations define DEA not as one invariant algorithm, but as a family of architectures in which specialization, discriminative selection, and structured condensation are used to avoid the brittleness of a single monolithic representation.