Forgery-Discriminative Units (FDUs)

• Forgery-Discriminative Units (FDUs) are specialized latent neurons identified through triadic fusion scoring that integrates activation, gradient, and probe weight for precise anomaly detection.
• They operationalize fixed or learned geometric anchors in deep neural representations to enforce class separability and invariance efficiently.
• Empirical studies reveal that FDUs achieve high accuracy in forgery detection benchmarks, supporting robust few-shot learning and effective transfer across tasks.

Discriminative Neural Anchors (DNA) constitute a set of methodological frameworks and exact computational recipes that operationalize the concept of fixed or learned anchor points in neural representations to induce class-separability, invariance, and explicit organization in the latent space. The DNA paradigm is instantiated in diverse applications such as deep discriminative representation learning, universal anomaly detection, and zero-shot learning, each grounded in rigorous formalization of anchor construction and loss-based alignment to these targets. Recent work encompasses both supervised settings, with fixed maximal-separation anchors shaping convolutional features, and unsupervised or transfer-based regimes, where anchors emerge via graph encoding or latent unit selection to localize discriminative knowledge.

1. Mathematical Definition and Anchor Construction Principles

DNA frameworks formalize “anchors” as target vectors in $\mathbb{R}^d$ assigned to each semantic class, feature, or compositional unit. In canonical supervised discriminative DNA (Hao et al., 2018), for $C$ classes, the anchor set $\{a_1,\ldots,a_C\}$ is prescribed statically, with:

• Unit-norm constraint: $\|a_c\|_2 = 1$ $\forall\,c$
• Maximal angular separation: For all $c\neq c'$, $\langle a_c, a_{c'} \rangle \leq \cos(\theta_M)$ for target margin $\theta_M$

Anchors are typically sampled as equiangular points on the unit hypersphere, achievable through (i) explicit geometric meshing or (ii) orthonormal basis selection when $d \geq C$. These anchors act as immutable targets for the deep representations, enforcing intra-class compactness and inter-class separability through direct feature-to-anchor pulling rather than pairwise or triplet-based contrast.

In zero-shot and transfer learning, discriminative anchors are synthesized by diffusion-based graph convolutional networks (GCN) operating over class–attribute bipartite graphs, yielding a set of anchor embeddings $\in \mathbb{R}^d$ with smoothed, disentangled geometry aligned to semantic regularities (Li et al., 2020). Anchors in transfer-based detection settings may also correspond to discovered latent neurons associated with class-distinction, identified by fusion metrics and pruning heuristics (Dou et al., 30 Jan 2026).

2. Loss Formulations and Optimization Algorithms

DNA models induce discriminative learning via anchor-based loss objectives. Two primary metric choices and their corresponding losses are formalized for CNNs (Hao et al., 2018):

• Euclidean Anchor Loss (E-NCM):

$$L_E(W) = -\frac{1}{N}\sum_{i=1}^{N}\log\frac{\exp\left(-\|f_W(x_i) - a_{y_i}\|_2\right)}{\sum_{c=1}^C \exp\left(-\|f_W(x_i) - a_c\|_2\right)}$$

• Cosine Anchor Loss (C-NCM):

$$L_C(W) = -\frac{1}{N}\sum_{i=1}^{N}\log\frac{\exp\left(-\left[1-\frac{f_W(x_i)^\top a_{y_i}}{\|f_W(x_i)\|}\right]\right)}{\sum_{c=1}^C \exp\left(-\left[1-\frac{f_W(x_i)^\top a_c}{\|f_W(x_i)\|}\right]\right)}$$

In these expressions, $f_W(x) \in \mathbb{R}^d$ is the deep feature for input $x$ and parameters $W$. Both losses feature a softmax over anchor distances, ensuring concurrent intra-class compactness and inter-class separability without the construction of example pairs or triplets, and are optimized by batch SGD.

In GCN-based DNA (DAGDA), anchors are learned via minimizing a diffusion-plus-regularization objective over the class–attribute graph adjacency. Anchor-alignment and distribution-regularization losses further enforce feature proximity to anchors and alignment of class-conditional distributions to attribute patternings, as detailed in the overall objective:

$$L = L_{\mathrm{cons}} + \lambda_1 L_{\mathrm{recons}} + \lambda_2 L_{\mathrm{reg}}$$

where each term encodes respective compatibility, reconstruction, and semantic relation losses (Li et al., 2020).

Universal anomaly detection regimes (e.g., AI-generated forgery detection) forego external labels, constructing anchors by identifying discriminative neurons (“forgery-discriminative units”, FDUs) via triadic fusion scoring, a composite product of activation magnitude, gradient sensitivity, and probe weight, thresholded by curvature-based truncation (Dou et al., 30 Jan 2026).

3. Algorithmic Instantiations and Computational Properties

Anchor-based DNA can be implemented with minimal modifications to conventional neural architectures. In supervised settings, training proceeds as follows (Hao et al., 2018):

1. Initialize quantized or geometric anchors, fixed for all epochs.
2. For each batch, compute feature–anchor distances for all samples and classes ($O(B\,C\,d)$), calculate softmax probabilities and cross-entropy loss.
3. Backpropagate only with respect to network parameters; anchors remain static.
4. Standard SGD optimization applies.

No anchor updates or pair/triplet mining occur. Modern DL frameworks permit efficient matrix-matrix implementation for distance computation.

In diffusion-GCN models for zero-shot learning, anchor embedding construction requires inversion/multiplication of normalized graph adjacency (or its $p$-step truncation), which may be computationally intensive for large graphs, but can leverage sparse algebra and batch computation. Distribution alignment for feature projection is realized via an auto-encoder with linear or nonlinear compatibility heads.

For universal anomaly detection, the critical stages include batchwise extraction of layer activations, computation of centroidal statistics, linear probe-based feature selection, and fusion-based neuron ranking. Curvature truncation (via the Kneedle algorithm) robustly selects anchor neurons without manual hyperparameter selection (Dou et al., 30 Jan 2026).

4. Empirical Performance and Comparative Results

Benchmark evaluation demonstrates that DNA frameworks outperform or match state-of-the-art approaches across diverse settings:

• Supervised classification (CNNs) (Hao et al., 2018):
  • On MNIST, CIFAR-10/100, anchor-based NCM losses yield error rates as low as $\sim0.47\%$ and $5.67\%$ (E-NCM) on MNIST and CIFAR-10+ respectively, outperforming softmax and L-softmax variants.
  • On CIFAR-100, E-NCM achieves $28.14\%$ error ($>1.8\%$ absolute improvement over softmax).
• Zero-Shot Learning (DAGDA) (Li et al., 2020):
  • On AWA2: $73.0\%$ mean class accuracy, exceeding SE-ZSL ($69.2\%$) and ZSKL ($70.5\%$).
  • Generalized ZSL: DAGDA achieves harmonic means exceeding all prior inductive methods on CUB and aPY.
• Generative Forgery Detection (Universal DNA) (Dou et al., 30 Jan 2026):
  • Outperforms MoLD and other baselines on ForenSynths/GenImage and on HIFI-Gen (new synthetic benchmark), with mean accuracy up to $97.2\%$ (cross-dataset) and up to $99.1\%$ on the most challenging generators.
  • Demonstrates extreme few-shot sufficiency, with core accuracy achieved using only $\sim5k$ samples.

A summary of representative results:

Benchmark	DNA Variant	Result	Baseline(s)
MNIST (supervised)	E-NCM	0.47% error	Softmax: 0.68%
CIFAR-100 (sup.)	E-NCM	28.14% error	Softmax: 37.26%
AWA2 (ZSL)	DAGDA	73.0% MCA	SE-ZSL: 69.2%
HIFI-Gen	DNA (forgery)	96.4% mean ACC	MoLD: 88.2%

5. Theoretical Properties and Interpretation

In anchor-based DNA, intra-class compactness and inter-class separability are explicitly regularized via direct loss minimization, bypassing implicit softmax organization or example mining. Anchors pre-arranged at large angles ensure a non-degenerate solution under the loss: collapse of all features to a single point is geometrically infeasible. In the Euclidean instance, the objective resembles a Gaussian mixture model with class means fixed to anchors and a uniform prior, though DNA does not perform generative modeling or mean learning—providing analytic tractability and stability.

GCN-based discriminative anchor generation introduces spectral smoothing that rectifies the entangled geometry of raw attribute spaces, producing anchor manifolds amenable to nearest-neighbor inference under strong distributional shift. Semantic relation regularization aligns feature–anchor distributions with original class–attribute topology, enhancing robustness in ZSL transfer (Li et al., 2020).

Universal DNA frameworks posit that pre-trained backbones already encode neurons ("sleeping" units) specialized for class-sensitive or anomaly-sensitive discrimination. The two-stage excavation protocol (layer localization, FDU identification) operationalizes this hypothesis, yielding robust, architecture- and pretraining-insensitive discriminative subspaces (Dou et al., 30 Jan 2026).

6. Limitations and Open Research Questions

Although discriminative neural anchor frameworks deliver efficient and competitive performance, limitations remain:

• Anchor rigidity: Fixed anchors (supervised DNA) cannot adapt to dataset or task-specific class distributions. Joint anchor-feature learning or dynamic anchor scheduling may address this.
• Packing constraints: For large $C$ in moderate $d$, deriving maximally separated anchors becomes intractable (a manifestation of the spherical code problem).
• Over-smoothing and generalization: In DAGDA, anchor graphs with few nodes or high attribute-class density may converge to over-smoothed, noisy anchor geometry, particularly in fine-grained or small-graph settings.
• Inductive assumptions: Universal DNA frameworks rely on assumed isotropy and transferability of latent features; theoretical understanding of the limits and structure of sleeping neuron manifolds is incomplete.
• Domain transfer: Existing evaluations focus on natural images; extension to medical, satellite, or other specialized domains remains an active area.

Potential extensions include angular margin penalties, graph attention for higher-order topology, anchor fine-tuning, and generative anchor-conditioned synthesis.

7. Connections to Broader Landscape and Methodological Impact

Discriminative neural anchors unify and generalize a range of concepts in representation learning:

• Bypassing the computationally intensive pair/triplet mining of contrastive/triplet losses via soft anchor-based cross-entropy.
• Enabling zero-shot and transfer regimes by graph-based construction of smooth, relational anchor spaces sensitive to unseen class semantics.
• Providing analytic interpretability in universal knowledge detection by precisely localizing the neurons most responsible for critical discrimination.

DNA frameworks have been integrated into diverse architectures (CNNs, ViTs, LLMs) and tasks (classification, detection, ZSL), underscoring the universality and flexibility of discriminative anchor construction (Hao et al., 2018, Li et al., 2020, Dou et al., 30 Jan 2026). Their role in producing robust, efficient, and interpretable representation geometries marks a significant direction in both classical and contemporary machine learning research.