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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Guard Me If You Know Me: Protecting Specific Face-Identity from Deepfakes
Authors: Kaiqing Lin, Zhiyuan Yan, Ke-Yue Zhang, Li Hao, Yue Zhou, Yuzhen Lin, Weixiang Li, Taiping Yao, Shouhong Ding, Bin Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our model outperforms existing methods in both detection and explanation. The code is available at https://github.com/KQL11/VIPGuard . 5 Experiments In this section, we present comprehensive experiments to evaluate the effectiveness of our method. |
| Researcher Affiliation | Collaboration | 1Guangdong Provincial Key Laboratory of Intelligent Information Processing, Shenzhen Key Laboratory of Media Security, and SZU-AFS Joint Innovation Center for AI Technology, Shenzhen University 2School of Electronic and Computer Engineering, Peking Univerisity, 3Tencent Youtu Lab |
| Pseudocode | No | The paper describes its methodology through three main stages of training (Face Attribute Learning, Identity Discrimination Learning, and User-Specific Customization) using mathematical equations (Eq. 3-7) and textual descriptions, but does not present a formal pseudocode or algorithm block. |
| Open Source Code | Yes | The code is available at https://github.com/KQL11/VIPGuard . |
| Open Datasets | Yes | To facilitate the evaluation of our method, we build a comprehensive identity-aware benchmark called VIPBench for personalized deepfake detection... We obtain all facial images from open-sourced datasets, including LAION-Face [87], Cross Face ID [64], and Face ID-6M [63]... Our VIPBench dataset is released under the Creative Commons Attribution-Non Commercial (CC BY-NC) license (more details can be seen in https: //creativecommons.org/licenses/by-nc/4.0/). |
| Dataset Splits | Yes | For each identity, 40-60 real images were collected, 20 reserved for testing in this benchmark DEval, and the remainder used to construct Dvip ID... The ratio of samples in Dgeneral ID is (Ir real It real Same ID) : (Ir real It real Diff ID) : (Ir real It fake Same ID) = 2 : 1 : 1, which ensures a balanced number of positive and negative samples... the ratio was adjusted to (Ir real It real Same ID) : (Ir real It real Diff ID) : (Ir real It fake Same ID) = 1 : 5 : 5 |
| Hardware Specification | No | To accommodate GPU memory limitations, the equivalent batch size was maintained at 72 for both Stage 1 and Stage 2 by applying gradient accumulation. All training was performed using mixed-precision computation within the open-source Swift5 framework. |
| Software Dependencies | Yes | This paper adopted the pre-trained Qwen-2.5-VL-7B model [1] as the backbone... Gemini API version in use: 2.5-pro-exp-03-25... GPT4o API version in use: GPT-4o-2024-08-06... All training was performed using mixed-precision computation within the open-source Swift5 framework. |
| Experiment Setup | Yes | Input images were resized to 448 × 448 when larger than this size. The model was optimized using the Adam optimizer with a cosine learning rate decay schedule, starting from an initial learning rate of 3e-5. To accommodate GPU memory limitations, the equivalent batch size was maintained at 72 for both Stage 1 and Stage 2 by applying gradient accumulation. In Stage 3, the effective batch size was reduced to 8 and the initial learning rate was set to 1. All training was performed using mixed-precision computation within the open-source Swift5 framework. The model was trained for 2 epochs in Stage 1, and for 1 epoch each in Stage 2 and Stage 3. |