Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
STD-FD: Spatio-Temporal Distribution Fitting Deviation for AIGC Forgery Identification
Authors: Hengrui Lou, Zunlei Feng, Jinsong Geng, Erteng Liu, Jie Lei, Lechao Cheng, Jie Song, Mingli Song, Yijun Bei
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that STD-FD effectively captures distribution patterns in AIGC-generated data, demonstrating strong robustness and generalizability while outperforming state-of-the-art (SOTA) methods on major datasets. |
| Researcher Affiliation | Academia | 1State Key Laboratory of Blockchain and Data Security, Zhejiang University 2School of Software Technology, Zhejiang University 3College of Computer Science, Zhejiang University of Technology 4School of Computer Science and Information Engineering, Hefei University of Technology 5Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security. Correspondence to: Yijun Bei <EMAIL>. |
| Pseudocode | Yes | In the appendix, we first present the detailed pseudocode of STD-FD (A) to facilitate a better understanding of the method. To further clarify the implementation process of STD-FD, we provide the core code for its key steps in addition to the open-source code. Specifically, algorithms 1 and 2 detail the generation and selection process of the critical spatio-temporal modeling factor, DFactor, in STD-FD. |
| Open Source Code | Yes | The source code is available at this link. |
| Open Datasets | Yes | We evaluate the proposed method using the Gen Image (Zhu et al., 2023) and Deepfacegen (Bei et al., 2024) datasets. |
| Dataset Splits | Yes | In our research, we follow the official dataset split, allocating 2,581,167 images for training and reserving the remaining 100,000 images for validation. Deep Face Gen is a facial forgery dataset that includes both video and image modalities. ... The dataset was split into training, validation, and test sets in a ratio of 7:1:2. |
| Hardware Specification | Yes | The experiments are conducted on two Ge Force RTX 4090 GPU (24GB VRAM). ... Experiments are conducted on a Silver 4310 CPU and an NVIDIA A40 GPU. |
| Software Dependencies | No | XGBoost (Chen & Guestrin, 2016) is employed as the classifier. In the identification feature construction phase, Euclidean distance and Dynamic Time Warping (DTW) are adopted to measure the similarity between the candidate discriminative factors and the optimal discriminative factors. For match rate-based methods, trend match rate (using sign matching) is applied, while Pearson correlation and template matching (NCC) are employed for correlation-based methods. |
| Experiment Setup | Yes | XGBoost (Chen & Guestrin, 2016) is employed as the classifier. In the identification feature construction phase, Euclidean distance and Dynamic Time Warping (DTW) are adopted to measure the similarity between the candidate discriminative factors and the optimal discriminative factors. For match rate-based methods, trend match rate (using sign matching) is applied, while Pearson correlation and template matching (NCC) are employed for correlation-based methods. ... To evaluate the effect of sampling timesteps on STD-FD, we conduct experiments on four challenging subsets of Deep Face Gen with sampling timesteps T set to 5, 10, 20, and 50. ... adversarial noise (with L2-norm strengths of [0.01, 0.03, 0.05]) was injected at each timestep of the reverse diffusion process (20 steps in total). ... We performed an ablation study varying the number of superpixel blocks K from the baseline setting K=10. |