COMPASS: Contextual Moment Prior Alignment for Generalizable AIGI Detection

Authors: Zhan Yang, Tiancheng Shen, Weijie Lyu, Shuo Chen, Bernard Ghanem, Yibo Yang, Ming-Hsuan Yang

Published: 2026-10-13 09:23:12+00:00

AI Summary

COMPASS addresses the challenge of generalizable AI-generated image (AIGI) and deepfake detection by introducing a data-aware subspace adaptation framework. It uses a small unlabeled Semantic Proxy Set to estimate target activation statistics, incorporating activation covariance into the decomposition of pre-trained Vision Foundation Model (VFM) weights. This approach preserves target-relevant representations as a stable semantic reference while a residual adapter learns forgery-related variations, achieving state-of-the-art performance across various generalization settings.

Abstract

Generalizable AI-generated image (AIGI) and deepfake detection remain challenging due to the asymmetry phenomenon, where adaptation to limited training distributions can degrade the transferable representations of Vision Foundation Models (VFMs). Recent subspace adaptation methods mitigate this issue by freezing high-magnitude principal components while adapting the residual subspace. However, such magnitude-based decomposition is data-agnostic and does not explicitly account for target-relevant semantics, potentially leading to semantic misalignment. We propose COMPASS (COntextual Moment Prior Alignment for Subspace Separation), a data-aware subspace adaptation framework that uses a small unlabeled Semantic Proxy Set to estimate target activation statistics. COMPASS incorporates activation covariance into the decomposition of pre-trained weights, preserving target-relevant representations as a stable semantic reference while the residual adapter learns forgery-related variations. The proxy calibration requires no gradient-based optimization. Extensive experiments demonstrate strong generalization across cross-dataset, cross-method, joint cross-dataset-method, and open-world cross-generator evaluations, achieving state-of-the-art performance on both Deepfake and AIGI detection benchmarks.

Key findings

COMPASS achieves state-of-the-art generalization in deepfake and AIGI detection, particularly outperforming magnitude-based methods in cross-method and joint cross-dataset-method evaluations. The approach is robust to proxy set selection, effective with small proxy sets, and generalizes across different VFM backbones. It also demonstrates superior robustness to real-world perturbations like JPEG compression and Gaussian blur.

Approach

COMPASS utilizes a small unlabeled Semantic Proxy Set to estimate target activation statistics. These statistics, specifically activation covariance, guide the Singular Value Decomposition (SVD) of pre-trained VFM weights. This covariance-guided decomposition separates weights into principal components (frozen as a stable semantic reference) and residual components (adapted for forgery-related variations).

Datasets

FF++ (c23), CDF-v1, CDF-v2, DFDC, DFDCP, UniFace, BleFace, MobSwap, e4s, FaceDan, FSGAN, InSwap, SimSwap, FOMM, FV2V, Wav2Lip, MRAA, OneShot, PiRender, TPSMM, LIA, DaGan, SadTalker, MCNet, HyperReenact, ProGAN, SDv1.4, R3GAN, StyleGAN-XL, FaceSwap, DALLE-3, FLUX1-dev, SDXL, Infinite-ID, CommunityAI (AIGIBench covering 25 generative models), ImageNet, COCO

Model(s)

CLIP ViT-L/14, BEiT-v2, SigLIP, DINOv3 (as VFM backbones)

Author countries

China, United States, Saudi Arabia

← Previous