DFA-Instruct
收藏资源简介:
DFA-Instruct数据集由马上消费金融有限公司和北京邮电大学的研究团队创建,旨在支持交互式深度伪造分析的研究。该数据集包含127,300张对齐的人脸图像和891,600个相关的问答对,涵盖了深度伪造检测、分类、伪影描述和自由对话等任务。数据集的构建过程通过GPT辅助生成,确保了数据的多样性和质量。该数据集的应用领域主要集中在信息安全、司法调查和个人隐私保护等领域,旨在通过多模态大语言模型的指令调优,提升深度伪造分析的交互性和准确性。
The DFA-Instruct dataset was developed by the research teams from Chongqing Ma Consumer Finance Co., Ltd. and Beijing University of Posts and Telecommunications, aiming to support research on interactive deepfake analysis. This dataset comprises 127,300 aligned facial images and 891,600 associated question-answer pairs, covering tasks such as deepfake detection, classification, artifact description, and free-form conversation. The dataset was built with the assistance of GPT to ensure its diversity and quality. Its application fields mainly focus on information security, judicial investigation, personal privacy protection and other related areas, with the purpose of enhancing the interactivity and accuracy of deepfake analysis through instruction tuning of multimodal large language models.




