GEPC: Group-Equivariant Posterior Consistency for Out-of-Distribution Detection in Diffusion Models

This paper introduces GEPC, a training-free method for out-of-distribution detection in diffusion models that leverages group-equivariant posterior consistency to identify anomalies by measuring score field transformation inconsistencies under symmetry groups, achieving competitive performance and interpretable results across image and SAR datasets.

The Gist

Imagine you have a very smart artist (a Diffusion Model) who has spent years learning to paint perfect landscapes, portraits, and cityscapes. This artist knows exactly how a "normal" tree, a "normal" face, or a "normal" building should look and feel.

Now, imagine someone hands the artist a picture of a flying toaster or a purple giraffe. The artist has never seen these things. How do you know the artist is confused?

Usually, we check if the artist's "confidence" is low. But sometimes, the artist might confidently say, "Oh yes, that's a very normal purple giraffe!" just because the colors are bright, even though the shape makes no sense.

This paper introduces a new way to catch the artist's confusion, called GEPC (Group-Equivariant Posterior Consistency). Here is how it works, using simple analogies:

1. The "Magic Mirror" Test

Imagine the artist is looking at a painting through a magic mirror.

• The Setup: You take a picture of a "normal" object (like a cat). You then take that picture, flip it upside down, rotate it, or shift it slightly (these are the "Group" actions).
• The Prediction: You ask the artist: "If I show you this rotated cat, what does the cat look like before it was rotated?"
• The Consistency Check:
  • For a Real Cat (In-Distribution): If you rotate the picture of a cat, the artist's brain should rotate the "answer" back perfectly. If you flip the picture, the answer flips back. The artist's internal logic is consistent. It's like a puzzle where all the pieces fit together no matter how you turn the box.
  • For a Flying Toaster (Out-of-Distribution): If you show the artist a flying toaster and rotate the picture, the artist's brain gets confused. It might try to rotate the answer in a way that doesn't match the original picture. The logic breaks. The pieces of the puzzle don't fit anymore.

2. Why This is Better Than Just "Looking"

Most old methods tried to detect weird things by looking at how "bright" or "dark" the artist's answer was (like checking if the artist is sweating).

• The Problem: A flying toaster might make the artist sweat just as much as a normal cat, or maybe not at all. The "sweat" (score magnitude) doesn't tell the whole story.
• The GEPC Solution: GEPC doesn't care about how much the artist is sweating. It cares about symmetry. It asks: "Does the artist's brain work the same way when I turn the world upside down?"
  • If the answer is Yes, it's probably a normal image.
  • If the answer is No (the logic breaks), it's likely a weird, out-of-distribution image.

3. The "Training-Free" Superpower

Usually, to teach a computer to spot weird things, you have to show it a million examples of "weird" things first. That takes a lot of time and data.

• GEPC is different: It doesn't need to be taught what a "flying toaster" looks like. It just uses the artist's existing knowledge of symmetry. It's like hiring a security guard who doesn't need a list of suspects; they just know that if someone walks through a door backwards, something is wrong.
• No Extra Cost: It doesn't require rebuilding the artist's brain or running complex extra calculations. It just asks the artist a few extra questions about rotated versions of the image.

4. Real-World Superpower: Finding Ships in the Ocean

The paper tested this on Radar Images (SAR), which are used to see ships in the ocean through clouds and at night.

• The Scene: The ocean is usually just "clutter" (waves, noise). This is the "Normal" data the artist learned.
• The Anomaly: A ship or a submarine is a "weird" object in that ocean.
• The Result: When GEPC looked at the radar images, it didn't just give a "Yes/No" answer. It drew a heat map. It highlighted exactly where the symmetry broke.
  • On empty water, the map was calm (consistent).
  • On the ship, the map lit up like a neon sign (inconsistent).
  • This means GEPC can tell you not just that there is a ship, but exactly where it is, even if the computer has never seen a ship before.

Summary

GEPC is a clever trick that checks if an AI's brain is "symmetrical."

• Normal things behave the same way when you flip or rotate them.
• Weird things break the rules of symmetry.
• By measuring how much the AI's logic "breaks" when you twist the image, GEPC can spot anomalies that other methods miss, without needing to be retrained. It's like catching a liar not by what they say, but by how they stumble when you ask the same question from a different angle.

Technical Summary

1. Problem Statement

Out-of-Distribution (OOD) detection is critical for deploying reliable machine learning systems. While Diffusion Models have emerged as powerful priors for OOD detection, existing methods primarily rely on:

• Score Magnitude: The norm of the score field ($\|s_\theta(x_t, t)\|$).
• Local Geometry: Curvature, covariance spectra, or trajectory energies along the reverse process.

Limitations of Current Approaches:

• They often ignore the equivariance properties inherent in diffusion models trained on data with symmetries (e.g., flips, rotations) using convolutional backbones.
• Many methods require computationally expensive operations, such as reverse trajectory sampling, Jacobian-vector products, or Hessian computations.
• They may fail to detect distribution shifts where the score magnitude remains similar, but the underlying symmetry structure is violated.

2. Methodology: GEPC

The authors propose Group-Equivariant Posterior Consistency (GEPC), a training-free probe that detects OOD inputs by measuring the consistency of the learned score field under group transformations.

Core Hypothesis

If a diffusion model is trained on In-Distribution (ID) data that is approximately invariant under a group $G$ (e.g., rotations, flips) and uses a convolutional backbone, the learned score field $s_\theta(x_t, t)$ should be approximately $G$-equivariant on ID samples.

• ID Behavior: Transforming a noisy input $x_t$ by $g \in G$ and transporting the predicted score back should yield the original score: $P_g^\top s_\theta(P_g x_t, t) \approx s_\theta(x_t, t)$.
• OOD Behavior: For OOD inputs (which violate learned symmetries or lie far from the ID manifold), this consistency breaks, resulting in a large residual.

The GEPC Algorithm

1. Forward Noising: Given an input $x_0$, sample a noisy version $x_t$ at a selected timestep $t$.
2. Group Transport: Apply a group transformation $g \in G$ to $x_t$ to get $P_g x_t$.
3. Score Evaluation & Transport Back:
   • Compute the score at the transformed point: $s_\theta(P_g x_t, t)$.
   • Transport this score back to the original coordinate frame: $\tilde{s} = P_g^\top s_\theta(P_g x_t, t)$.
4. Residual Calculation: Compute the equivariance residual field:
   $$r_t(x_t, g) = \tilde{s} - s_\theta(x_t, t)$$
5. Aggregation:
   • Compute the squared $L_2$ norm of the residual: $R_t = \|r_t\|_2^2$.
   • Pool spatially and average over the group $G$ and multiple noise samples.
   • Normalize by the pooled energy of the original score to ensure stability.
   • Aggregate across a selected set of timesteps $T$ (weighted by stability).
6. Calibration: The final scalar score is calibrated using only ID data (via KDE, Z-score, or Mahalanobis distance) to determine the OOD threshold.

Key Theoretical Insights

• Population Bounds: The authors derive theoretical upper bounds for ID residuals and lower bounds for OOD residuals. They show that the expected residual is driven by the equivariance-breaking functional $B(G)(p_t)$ and the score approximation error.
• Cross-Backbone Robustness: Even if the diffusion backbone is trained on a different source distribution (e.g., LSUN images applied to SAR radar), GEPC remains effective. Theoretically, the residual scales with the distance of the test sample from the source manifold, provided the score field is Lipschitz continuous.
• Sensitivity to Mean Shifts: Unlike score magnitude (which is invariant to mean shifts in Gaussian distributions), GEPC is highly sensitive to mean shifts and anisotropy, making it a more robust OOD signal.

3. Key Contributions

1. Novel OOD Metric: Introduction of GEPC, a training-free method that leverages group-equivariance breaking in diffusion score fields rather than score magnitude or trajectory energy.
2. Theoretical Guarantees: Derivation of population-level bounds showing that GEPC separates ID and OOD based on symmetry violations and distance-to-manifold, without requiring Jacobian computations.
3. Practical Efficiency:
   • Requires no fine-tuning, no architectural changes, and no Jacobian/Hessian evaluations.
   • Computational cost is comparable to simple score-norm baselines (requiring only forward passes) but significantly cheaper than trajectory-based methods.
   • Produces interpretable heatmaps highlighting exactly where symmetry breaking occurs (spatial localization).
4. Comprehensive Evaluation:
   • Validated on standard benchmarks (CIFAR-10, SVHN, CelebA) showing competitive or superior AUROC compared to state-of-the-art diffusion baselines (e.g., SCOPED, DiffPath).
   • Demonstrated success in a cross-domain, high-resolution setting: Detecting ships/anomalies in Synthetic Aperture Radar (SAR) imagery using a backbone trained on natural images (LSUN), a task where traditional methods often struggle.

4. Experimental Results

• Benchmark Performance (32x32): On 9 ID/OOD pairs using a single CelebA-trained backbone, GEPC achieved an average AUROC of 0.908, outperforming or matching methods like SCOPED (0.892) and DiffPath (0.918) while requiring significantly fewer function evaluations (16 F + 0 J vs. 1000 F for NLL).
• SAR Imagery (256x256): Applied to HRSID and SSDD datasets (ship detection in sea clutter).
  • GEPC achieved 0.853 AUROC on intra-dataset ship detection and 1.000 AUROC on cross-dataset (SSDD) detection.
  • Crucially, it generated equivariance-breaking maps that visually localized ships and wakes against homogeneous sea clutter, providing interpretability that scalar scores lack.
• Ablation Studies:
  • Performance is stable across different group elements (flips, rotations, shifts), indicating the method captures a global symmetry-breaking effect rather than relying on a single artifact.
  • Timestep selection based on ID-only stability (Coefficient of Variation) effectively identifies the most informative noise levels for detection.

5. Significance and Impact

• New Paradigm for OOD: Shifts the focus from "how large is the score?" to "how consistent is the score's behavior under symmetry?" This leverages the inductive biases of the model architecture itself as a diagnostic tool.
• Computational Efficiency: By avoiding reverse trajectories and Jacobian calculations, GEPC makes diffusion-based OOD detection feasible for real-time or resource-constrained applications.
• Interpretability: The ability to generate spatial heatmaps of equivariance breaking is a major advantage for safety-critical domains (e.g., radar, medical imaging), allowing operators to see why an input is flagged as OOD.
• Cross-Domain Applicability: The success on SAR imagery using a natural image backbone suggests GEPC is robust to domain shifts, making it a strong candidate for "foundation model" reuse in diverse sensing modalities.

In summary, GEPC provides a theoretically grounded, computationally lightweight, and interpretable framework for OOD detection that exploits the geometric symmetries inherent in diffusion models, outperforming existing methods in both standard benchmarks and challenging cross-domain scenarios.