Generalizable Detection of AI Generated Images with Large Models and Fuzzy Decision Tree

Fei Wu; Guanghao Ding; Zijian Niu; Zhenrui Wang; Lei Yang; Zhuosheng Zhang; Shilin Wang

cs.CVMar 30, 2026

Generalizable Detection of AI Generated Images with Large Models and Fuzzy Decision Tree

Authors: Fei Wu, Guanghao Ding, Zijian Niu, Zhenrui Wang, Lei Yang, Zhuosheng Zhang, Shilin Wang

Organizations: Shanghai Jiao Tong University

Abstract

The malicious use and widespread dissemination of AI-generated images pose a serious threat to the authenticity of digital content. Existing detection methods exploit low-level artifacts left by common manipulation steps within the generation pipeline, but they often lack generalization due to model-specific overfitting. Recently, researchers have resorted to Multimodal Large Language Models (MLLMs) for AIGC detection, leveraging their high-level semantic reasoning and broad generalization capabilities. While promising, MLLMs lack the fine-grained perceptual sensitivity to subtle generation artifacts, making them inadequate as standalone detectors. To address this issue, we propose a novel AI-generated image detection framework that synergistically integrates lightweight artifact-aware detectors with MLLMs via a fuzzy decision tree. The decision tree treats the outputs of basic detectors as fuzzy membership values, enabling adaptive fusion of complementary cues from semantic and perceptual perspectives. Extensive experiments demonstrate that the proposed method achieves state-of-the-art accuracy and strong generalization across diverse generative models.

Figures & tables

Fig. 1: An example where an MLLM fails to detect a fake image due to its insensitivity to fine-grained visual artifacts.

Fig. 2: An overview of our proposed method, which integrates lightweight detectors and multimodal large language models via a fuzzy decision tree.

Model	Real	GAN-based	DM-based	In-the-Wild	Fake	Overall
UnivFD [ 13 ]	95.03	66.18	14.60	23.99	28.53	61.81
PatchCraft [ 22 ]	72.11	99.23	58.14	5.58	51.55	61.84
NPR [ 15 ]	91.19	94.99	81.12	7.53	62.45	76.84
FreqNet [ 14 ]	85.03	69.00	50.89	14.57	44.11	64.59
DRCT [ 2 ]	38.74	83.88	90.18	89.00	88.47	63.58
AIDE [ 18 ]	81.31	67.09	82.34	36.11	65.43	73.37

TABLE I: Detection accuracy (%) of lightweight detectors on aigibench across different data categories.

Fig. 3: An example of the structured prompt used for MLLMs

Fig. 4: Accuracy of Qwen3-VL-32B-Instruct with different prompts. Each cell is subdivided into four subcells, representing four output prompts.

Model	AIGIBench	GenImage	AIGCDet	WildRF	FakeBench	Chameleon	Overall	Avg.	Std.	Robustness
Lightweight Detectors
NPR † [ 15 ]	76.84	93.71	93.90	67.52	61.37	59.25	84.27	75.43	14.14	71.35
DRCT † [ 2 ]	63.58	89.69	84.08	53.46	85.47	53.86	74.77	71.69	15.18	65.89
AIDE † [ 18 ]	73.37	87.40	83.07	67.28	65.37	62.61	78.45	73.18	9.20	70.41
UnivFD [ 13 ]	61.81	69.71	74.94	57.61	71.65	57.40	67.27	65.52	6.91	57.90
PatchCraft [ 22 ]	61.84	67.49	77.92	65.16	50.37	58.08	67.56	63.48	8.48	66.68

TABLE II: Comparison of baselines and the fuzzy decision tree. Here,

\dagger

denotes models used as base components for the ensemble strategies. Bold and underline indicate the best and second-best performance in each column, respectively. Overall denotes the sample-weighted accuracy, while avg. and std. are computed across benchmarks. Robustness reports the mean accuracy under blur/jpeg/resize on aigibench-dev. aigibench is evaluated on the remaining test set excluding aigibench-dev.

Lightweight	MLLMs	Overall	Avg.	Std.	Robustness
✓		87.32	81.11	11.75	79.99
	✓	77.83	82.08	6.22	84.69
✓	✓	90.14	90.04	5.76	87.40

TABLE III: Ablation study of the proposed fuzzy decision tree.

Fig. 5: Visualized structure of the fuzzy decision tree with inference paths for two representative fake image examples. Each internal node is parameterized by a fuzzy predicate $(\mathcal{S},\phi,\tau)$ , where a selected subset of detectors $\mathcal{S}$ is fused via an ensemble operator $\phi$ and compared against a threshold $\tau$ .

AI-Generated Image Detection

AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.

See all 186 papers

Perception, Verdict, and Evolution: Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection
Jun 25, 2026Yangjun Wu, Keyu Yan, Yu Liu +5LLM Self-Refinement AI-Generated Image Detection
Zhejiang University, Zhejiang, China · Alibaba Group
Reduce the Artifacts Bias for More Generalizable AI-Generated Image Detection
May 14, 2026Yiheng Li, Yang Yang, Zichang Tan +3Synthetic Data Augmentation Mixture of Experts
School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · MAIS, Institute of Automation, Chinese Academy of Sciences, Beijing, China · Sangfor Technologies Inc. +4
Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection
May 24, 2026Zijie Cao, Weijie Tu, Yao Xiao +3Synthetic Data Augmentation AI-Generated Image Detection
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Australian National University, Canberra, Australia · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China +1

Jun 25, 2026J/K move · Enter open · S save

Perception, Verdict, and Evolution: Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection

Yangjun Wu, Keyu Yan, Yu Liu, Jingren Zhou+4

Zhejiang University, Zhejiang, China · Alibaba Group

LLM Self-Refinement AI-Generated Image Detection

Open paper AskPDF arXiv

The rapid advancement of generative models presents a significant challenge to existing deepfake detection methods, particularly given the widespread dissemination of highly realistic AI-generated images. Although Multimodal Large Language Models (MLLMs) show strong potential for this task, existing approaches suffer from two key limitations: insufficient sensitivity to fine-grained forensic artifacts and reliance on static synthetic supervision from frontier models, leading to limited flexibility and high-cost. To address these issues, we propose ForeAgent, an agentic forensics framework for AI-generated image detection with iterative self-evolution. First, ForeAgent adopts a Perception-Verdict architecture that aggregates multi-view cues spanning semantic, spatial, and frequency-domain features, and leverages an MLLM as a verdict module to fuse these signals for a logical-grounded verdict. Second, to enable continual self-improvement, we introduce a Hindsight-Driven Self-Refining strategy following a Sampling-Reflection-Evolution paradigm. The agent performs inference rollouts on training instances. Guided by ground-truth labels as hindsight, it reflects on failure cases and low-quality reasoning trajectories to regenerate higher-quality reasoning traces. These synthesized samples are then strictly filtered through a dual-expert quality gating module. ForeAgent continuously evolves via fine-tuning on self-curated high-quality samples. Extensive experiments demonstrate that ForeAgent achieves state-of-the-art performance on the Chameleon benchmark, reaching 82.18% accuracy (+16.41% over AIDE), and achieves 93.3% mean accuracy on AIGCDetect-Benchmark across 16 generators. In addition, external evaluation shows that ForeAgent produces more consistent and causally grounded reasoning compared to GPT-5 and GPT-5-mini.

Abstract

AI-Generated Image Detection

Explore similar work

Perception, Verdict, and Evolution: Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection

Reduce the Artifacts Bias for More Generalizable AI-Generated Image Detection

Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection