DRIFT: From Robustness Gaps to Invariance Manifolds for AI-Generated Image Detection

2026-06-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Image Processing · Depth: Expert, extended

Summary

DRIFT is a novel AI-generated image detection framework developed by Samsung Research Institute that addresses limitations of existing training-free methods by learning a structured invariance manifold of real images. It utilizes a frozen DINOv2 ViT-B/14 backbone with two lightweight projection heads (robust and fragile) to decompose representation space. The robust subspace suppresses variations from physically plausible transformations, while the fragile subspace retains sensitivity to edit-like perturbations. A structured ordering margin of γ=0.3 enforces hierarchical separation, enabling detection as a margin-violation test. The framework incorporates an EMA teacher with momentum 0.996 and a reconstruction anchor (weight 0.1) to stabilize training on real-only data. Experiments show strong open-world generalization, achieving a mean ACC/AP of approximately 97.8/99.8 on ForenSynth, consistently high accuracy on Diffusion-6cls, and 93.2% ACC / 92.0% AP on Gemini and 94.8% ACC / 95.0% AP on ChatGPT for PromptWorld-1K. It also provides interpretable patch-wise localization heatmaps.

Key takeaway

For Machine Learning Engineers developing robust AI-generated image detectors, you should consider adopting a structured invariance learning approach. This method, which explicitly models real-image manifolds using robust and fragile representation subspaces, offers superior open-world generalization compared to fixed robustness gap techniques. Implement an EMA teacher and reconstruction anchor to stabilize training on real-only datasets, and utilize patch-wise drift maps for both detection and interpretable localization of synthetic content.

Key insights

AI-generated image detection improves by learning a structured invariance manifold of real images using robust and fragile representation subspaces.

Principles

• Decompose representation space into robust and fragile subspaces.
• Enforce hierarchical ordering between physical invariance and edit variability.
• Stabilize one-class invariance learning with EMA and reconstruction.

Method

Train projection heads on a frozen VFM using real-only data, enforcing robust invariance, fragile sensitivity, and an ordering margin with EMA and reconstruction losses. Detect fakes via margin violation.

In practice

• Use patch-wise drift maps for interpretable AI-generated image localization.
• Aggregate patch scores with Top-k median for robust global detection.

Topics

• AI-Generated Image Detection
• Invariance Learning
• Representation Learning
• Vision Foundation Models
• Digital Forensics
• Image Authenticity

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• DRIFT: From Robustness Gaps to Invariance Manifolds for AI-Generated Image Detection — AI-generated image detection can be framed as learning a real-image invariance manifold to identify margin violations.
• The Matching Principle: A Geometric Theory of Loss Functions for Nuisance-Robust Representation Learning — Robustness techniques converge on estimating and regularizing encoder Jacobians against label-preserving deployment nuisance covariance.
• Robust Representation Learning through Explicit Environment Modeling — Explicitly modeling and marginalizing environmental variation yields more robust OOD generalization than invariance-seeking methods.
• How Useful is Causal Invariance for Domain Adaptation in Finite-Sample Settings? — Causal invariance improves supervised domain adaptation in finite-sample settings when target-risk margins between candidate models are sufficiently large.
• Evaluating the Representation Space of Diffusion Models via Self-Supervised Principles — A new framework uses self-supervised principles and a Fisher-based metric (ICR) to evaluate diffusion model representations and detect memorization.
• JA-SIREN: Deterministic Initialization for Sinusoidal Networks via Spectral Matching — Deterministic initialization via spectral matching for sinusoidal networks significantly enhances INR reproducibility and performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.