TRiGS: A Multifaceted Research Framework

• TRiGS is a term that denotes distinct, rigorously defined frameworks across deep learning security, dynamic scene reconstruction, 3D generation, and algebraic graph theory.
• It includes a Trojan detection method using inversion-driven gradient signatures and a 4D Gaussian Splatting approach that models continuous rigid-body motion with Bézier residuals.
• These techniques offer robust performance improvements and scalability, advancing both theoretical graph classifications and practical applications in computer vision and graphics.

TRiGS refers to multiple, rigorously defined concepts across disparate research communities, each notable for its technical significance and impact. The term may denote: (1) "TRojan Identification from Gradient-based Signatures," a generic framework for detecting Trojaned deep learning models using inversion-driven gradient signatures (Hussein et al., 2023); (2) "Temporal Rigid-Body Motion for Scalable 4D Gaussian Splatting," a dynamic scene representation for scalable and memory-efficient 4D rendering with continuous rigid-body motion modeling (Yeom et al., 1 Apr 2026); (3) the triplane-based Gaussian splatting field representation for 3D generative modeling (DirectTriGS) (Ju et al., 10 Mar 2025); and (4) "Triply-transitive Strongly-Regular Graphs" in algebraic graph theory, which are classified by the structure of associated Terwilliger and centralizer algebras (Herman et al., 18 Jul 2025). Each usage is technically disjoint and is described below in turn.

1. TRojan Identification from Gradient-based Signatures (TRIGS)

Formalization and Signature Construction

TRIGS provides a data-driven, white-box detector for identifying Trojaned classifiers $f:\mathbb{R}^{C\times H\times W}\to\mathbb{R}^K$. It constructs a per-model signature $S$ by optimizing inputs $x$ to minimize and maximize each logit $f_j(x)$. For $M=2K$, let

$$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$

with $L_{L2}(x) = \|x\|_2$ and $L_{TV}(x)$ as isotropic total variation. For each $i$, solve

$$a_i = \arg\min_{x\in\mathbb{R}^{C\times H\times W}}\;\; L_i(f(x)) + \lambda_{L2}\cdot R_{L2}(x) + \lambda_{TV}\cdot R_{TV}(x)$$

and concatenate $S$0. Regularization parameters are set (e.g.) as $S$1, $S$2.

Training and Input Protocols

TRIGS requires model weights but not any poisoned data to generate signatures. For downstream detection, a shadow set of benign and Trojan models (often as few as 250–1000) is assembled, each trained from limited clean data (4–10% Tiny-ImageNet). No knowledge of the specific attacked architecture, dataset, or trojaning strategy is assumed for the detector.

Downstream Classifier and Feature Aggregation

Signatures $S$3 are converted into feature tensors $S$4 by computing, at each pixel, 11 summary statistics (min, max, mean, std, quantiles, histogram bins) over the 2K signature channels, decomposed into minimization, maximization, and joint maps. A ResNeXt-50 binary CNN, modified to accept $S$5 or $S$6 input channels, is then trained with cross-entropy loss and Adam.

Theoretical Rationale

TRIGS exploits trigger-logit coupling: in Trojaned models, the adversarial trigger is so tightly bound to a target class that optimizing any logit $S$7 causes trigger motifs to emerge, independently of $S$8. Thus both minimization and maximization inversions reveal the backdoor. Clean models, lacking such universal patterns, generate merely class-typical textures.

Experimental Results and Comparisons

On CIFAR10 (VGG), Tiny-ImageNet (ResNet10), and ImageNet (ViT-B-16), TRIGS outperforms Universal Litmus Patterns and k-Arm Optimization for ROC-AUC (CIFAR10: $S$9, Tiny-ImageNet: $x$0, ImageNet: $x$1 using pixel-statistics). Activation minimization alone was often optimal, a novel empirical finding (Hussein et al., 2023).

Method	CIFAR10	Tiny-ImageNet	ImageNet
TRIGS-Stats	0.99 ± 0.003	1.00 ± 0.001	0.84 ± 0.05
ULP (baseline)	0.64 ± 0.06	0.74 ± 0.08	0.58
k-Arm	0.48 ± 0.01	0.56 ± 0.12	0.50 ± 0.07

TRIGS’s performance is robust to reduced clean data, few shadow models, and model architecture mismatch.

2. Temporal Rigid-Body Motion for Scalable 4D Gaussian Splatting (TRiGS)

Problem Setting and Motivation

TRiGS addresses memory and scalability limitations in 4D Gaussian Splatting (4DGS) for dynamic scene reconstruction by modeling each Gaussian primitive's trajectory as a continuous, unified SE(3) rigid-body motion, refined by Bézier residuals and stabilized with learnable local anchors (Yeom et al., 1 Apr 2026).

Motion and Temporal Representation

A scene is a fixed set $x$2 of Gaussian primitives with time-dependent position $x$3, covariance $x$4, color, and opacity. Motion is parameterized as

$x$5

where $x$6 is a time-varying twist in $x$7. Bézier residuals $x$8 allow for nonconstant acceleration.

Local Anchors and Temporal Identity

Each primitive has a local anchor $x$9 projected orthogonally to the instantaneous rotation axis, ensuring rigid, stable, part-wise motion. This anchoring prevents degeneracy and enables temporally consistent tracking over hundreds or thousands of frames, without needing to spawn or eliminate primitives mid-sequence.

Optimization and Learning

All parameters (means, covariances, opacities, motion, anchors) are learned via a composite loss: $f_j(x)$0 where $f_j(x)$1 is L2 photometric, $f_j(x)$2 enforces local motion coherence via color affinity-weighted differences, and $f_j(x)$3 penalizes excessive Bézier acceleration.

Results and Scalability

On the SelfCap and Neural 3D Video datasets (upto 1200 frames), TRiGS achieves higher fidelity (PSNR up to $f_j(x)$4) and sharper results than prior piecewise-linear or deformation-based models, with constant primitive budget and memory ($f_j(x)$5 MB) (Yeom et al., 1 Apr 2026).

Method	Frames	PSNR	Gaussians	Memory
TRiGS	1200	26.05	0.5M	160 MB
FreeTimeGS	1200	25.41	4M	977 MB

TRiGS is the first 4D GS framework to combine continuous, coupled rigid-body motion, Bézier residuals, and local anchors to achieve scalable, long-sequence, memory-efficient reconstruction.

3. Triplane-based Gaussian Splatting for 3D Generation (DirectTriGS)

Representation: Triplane Field Encoding

DirectTriGS encodes a Gaussian-splatting point cloud,

$f_j(x)$6

into three dense 2D feature planes (F), mapping any $f_j(x)$7 to triplane features via bilinear interpolation. This enables the representation of GS as a continuous image-like field (Ju et al., 10 Mar 2025).

Differentiable TriRenderer and Decoding

The TriRenderer provides a fully differentiable pipeline:

1. Geometry branch: C_geo channels predict SDF $f_j(x)$8 mesh extraction (FlexiCubes).
2. Surface sampling: Elliptical Gaussian primitives are sampled along the mesh surface.
3. Attribute branch: C_app channels decode to Gaussian attributes and splat rendering.

Rendering loss combines silhouette ($f_j(x)$9), RGB/SSIM ($M=2K$0), and perceptual ($M=2K$1) terms.

Compression and Generation via VAE and Diffusion

A triplane VAE compresses F into a latent code, with separated geometry and appearance branches. Latent diffusion models (LDMs) support text-conditioned 3D object generation:

• Stage 1 generates geometry latent $M=2K$2,
• Stage 2 generates appearance latent $M=2K$3 conditioned on $M=2K$4 and text.

The pipeline thus enables text-to-3D synthesis via direct sampling in the compressed triplane space, followed by TriRenderer-based GS field decoding.

Summary of Core Equations

• Triplane lookup: $M=2K$5
• Renderer: $M=2K$6
• Loss composite: $M=2K$7
• Diffusion: $M=2K$8

This framework bypasses per-object optimization and achieves high-quality 3D GS-based generation from text (Ju et al., 10 Mar 2025).

4. Triply-Transitive Strongly-Regular Graphs ("TRiGS") in Algebraic Graph Theory

Definitions and Algebraic Structure

A finite graph $M=2K$9 is strongly regular with parameters $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$0 if it is $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$1-regular, each pair of adjacent vertices shares $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$2 neighbors, and each non-adjacent pair shares $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$3 neighbors. The Terwilliger algebra $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$4 at vertex $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$5 is generated by adjacency and diagonal idempotent matrices: $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$6 where $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$7 is the adjacency matrix and $$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$8 its complement (Herman et al., 18 Jul 2025).

$$L_i(x) = \begin{cases} f_i(x) & 1\leq i \leq K\ -f_{i-K}(x) & K+1 \leq i \leq 2K \end{cases}$$9 is triply transitive if $L_{L2}(x) = \|x\|_2$0 (the centralizer algebra of the vertex stabilizer) for any $L_{L2}(x) = \|x\|_2$1.

Classification Theorem

Except for two geometric infinite families—collinearity graphs of $L_{L2}(x) = \|x\|_2$2 and affine polar graphs $L_{L2}(x) = \|x\|_2$3—triply-transitive strongly-regular graphs are classified as follows:

Family/Graph	Parameters	Automorphism Group
Complete $L_{L2}(x) = \|x\|_2$4-partite ($L_{L2}(x) = \|x\|_2$5)	$L_{L2}(x) = \|x\|_2$6	$L_{L2}(x) = \|x\|_2$7
5-cycle $L_{L2}(x) = \|x\|_2$8	$L_{L2}(x) = \|x\|_2$9	$L_{TV}(x)$0
McLaughlin	$L_{TV}(x)$1	$L_{TV}(x)$2
Higman–Sims	$L_{TV}(x)$3	$L_{TV}(x)$4
Paley($L_{TV}(x)$5), Paley($L_{TV}(x)$6)	$L_{TV}(x)$7, etc.	$L_{TV}(x)$8
Grid $L_{TV}(x)$9	$i$0	$i$1

The proof leverages the classification of rank-3 permutation groups and Krein-parameter bounds. Only the stated graphs and the two geometric families meet the condition $i$2 (Herman et al., 18 Jul 2025).

Conjectural Families

The infinite geometric families expected to be triply–transitive are:

• Collinearity graphs of the orthogonal polar space $i$3.
• Affine polar graphs $i$4.

Both remain open cases in the full classification.

5. Distinctions and Context

The term "TRiGS" encompasses conceptually unrelated frameworks:

• Trojan identification via gradient signatures is foundational in model security and interpretability (Hussein et al., 2023).
• Temporal rigid-body 4D Gaussian splatting addresses practical challenges in scalable, long-sequence dynamic scene modeling (Yeom et al., 1 Apr 2026).
• Triplane-based 3D generation leverages the triplane field and Gaussian splatting for high-fidelity generative modeling (Ju et al., 10 Mar 2025).
• Triply-transitive SRGs arise in algebraic combinatorics and permutation group theory (Herman et al., 18 Jul 2025).

Each usage is context-specific, and researchers should disambiguate the meaning based on discipline and cited work.