Do Transformations Reveal the Truth? Generative Residual Learning for Generalized AI-Generated Image Detection

Authors: Kutub Uddin, Nusrat Tasnim, Awais Khan, Mohammad Umar Farooq, Khalid Malik

Published: 2026-07-09 16:36:54+00:00

AI Summary

This paper introduces GenRes and GenRes++, novel frameworks for AI-generated image (AIGI) detection that leverage generative residual learning. GenRes models fine-grained relational features between original and transformed images via a neural tensor network, enhancing generalization across diverse generative methods. GenRes++ further improves this by incorporating a learnable attention mechanism to aggregate features from multiple transformed samples, focusing on the most informative cues for robust detection of AI-generated content, especially from unseen methods.

Abstract

The rapid advancement of generative AI has enabled the creation of highly realistic deepfake media, posing significant threats, including misinformation, digital identity theft, fraud, and manipulation of public opinion. AI-generated image (AIGI) detection is reliably challenging due to the diversity of generative methods and the subtle artifacts they leave behind. In this work, we propose GenRes, a novel framework for generative residual learning via a neural tensor network, which models fine-grained relational features between original and transformed samples to enhance generalization. To address scenarios involving multiple generative transformations, we introduce GenRes++, which employs a learnable attention mechanism to aggregate relational features across multiple transformed samples and enables the model to focus on the most informative cues. Both models leverage PE-Core as a feature extractor, providing generalized and semantically rich embeddings that improve cross-domain performance and enable the detection of AIGI generated by unseen methods. Comprehensive experiments on multiple benchmark datasets demonstrate that the proposed GenRes++ approach outperforms existing methods.

Key findings

GenRes++ achieved state-of-the-art performance with 95.7% mACC and 99.1% mAP across 19 unseen generative models, outperforming existing methods. The Neural Tensor Network, along with the diversity and number (N=5) of transformations, was found to be crucial for this high performance and cross-generator generalization. The framework demonstrates stable performance across GAN, other, and diffusion model families, highlighting its robustness.

Approach

The approach involves encoding an original image and its multiple generatively transformed variants using a shared PE-Core vision encoder with LoRA fine-tuning. For GenRes++, a Cross-Attention Aggregation (CCA) module then dynamically weights and combines these transformed features. Finally, a Neural Tensor Network (NTN) models the bilinear relational interactions between the original image's embedding and the aggregated transformed features, followed by a binary classification head to determine if the image is real or AI-generated.

Datasets

UniversalFakeDetect

Model(s)

GenRes, GenRes++, PE-Core-G14-448 (as backbone), LoRA, Neural Tensor Network (NTN)

Author countries

USA, South Korea

← Previous