New AI Detector Splits ‘Fake’ Into Subclasses to Boost Generalisation

Researchers have proposed a new framework for detecting AI-generated images and videos that aims to remain reliable when tested on content types it has never seen before. The work, published as an open-access original article in Complex & Intelligent Systems, introduces a Semantic Generalisable Deepfake Detection (SGDFD) model built around a sparsely gated Mixture of ORthogonal Decomposition (MoORD) architecture.

The authors argue that the generalisability of AI-generated content (AIGC) detection can be decomposed into two tightly coupled subspaces: a semantic subspace and a forgery subspace.

A second limitation the researchers identify concerns how existing methods label their training data. Most approaches consolidate all forged images and videos, regardless of which deepfake generation technique produced them, into a single monolithic “fake” class. The authors contend that this design choice overlooks the distinctive characteristics and substantial variability associated with different generation methods.

To address both issues, the SGDFD framework combines two technical components. First, it employs a parameter-efficient fine-tuning strategy based on orthogonal decomposition, which the authors report enhances the model’s generalisation capability in AIGC detection. Second, the MoORD architecture is designed to improve semantic generalisability, allowing the detector to perform across varying content categories.

The framework also learns a product-manifold representation that decomposes the single “fake” class into multiple sub-classes, each corresponding to a distinct deepfake generation method. In principle, this finer-grained structure lets the model distinguish between different forgery techniques rather than treating all synthetic media as an undifferentiated category.

The reported evaluation is broad in scope. The authors describe extensive experiments conducted on 10 real-world AIGC detection datasets under 3 evaluation protocols. According to the abstract, the proposed SGDFD model achieves state-of-the-art generalisation performance in AIGC detection across both known and unknown semantic categories.

The distinction between known and unknown semantic categories is central to the paper’s claims.

The article was published on 20 October 2026 by Qian Liu, Sonja Schimmler and Manfred Hauswirth, and is available open access.

Subject of Research: A deepfake detection framework for AI-generated images and videos that separates semantic and forgery subspaces and distinguishes between different generation methods.

Article Title: Semantic generalisable multimodal AI generated content detection via mixture of orthogonal decomposition and product manifold disentanglement

Article References: Liu, Q., Schimmler, S., & Hauswirth, M. (2026). Semantic generalisable multimodal AI generated content detection via mixture of orthogonal decomposition and product manifold disentanglement. Complex & Intelligent Systems, 12(12), Article 273. https://doi.org/10.1007/s40747-026-02536-5

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02536-5

Keywords: AI-generated content detection, deepfake detection, generalisability, mixture of orthogonal decomposition, product manifold, parameter-efficient fine-tuning, multimodal detection, semantic subspace, forgery subspace, machine learning, Semantic, generalisable