汕头大学学报（自然科学版）

随着人工智能（artificial intelligence,AI）生成动画在娱乐、教育等领域的广泛应用，其暴力、血腥等敏感内容对青少年身心健康及网络生态安全构成严重威胁.为解决传统敏感内容识别方法依赖人工特征、难以适应动画多样性与动态性的问题，研究通过融合红绿蓝（red green blue,RGB）图像特征与光流特征，提出了一种基于双模3D卷积神经网络的AI动画敏感内容识别系统.实验结果表明，在Sakuga-42M数据集上，双模3D卷积神经网络在训练集的平均准确率为97.09%，高风险场景召回率97.21%，远高于传统卷积神经网络的86.04%和86.57%.实际应用中，AI动画敏感内容识别系统的人工复核率仅5.27%，更新周期为3.5天.可见，该系统有着较高的敏感内容识别准确率和较低的人工复核成本.研究所提方法为AI动画内容安全审核提供了高精度、低成本的解决方案，可有效保障网络环境的健康安全.

With the widespread application of artificial intelligence(AI)generated animation in entertainment, education, and other fields, its sensitive content such as violence and gore poses a serious threat to the physical and mental health of young people and the security of the online ecosystem. To solve the problem of traditional sensitive content recognition methods relying on artificial features and being difficult to adapt to the diversity and dynamics of animation, a dual-mode 3D convolutional neural network-based AI animation sensitive content recognition system is proposed by integrating RGB image features and optical flow features. The experimental results show that on the Sakuga-42M dataset, the dual-mode 3D convolutional neural network has an average accuracy of 97.09% in the training set and a recall rate of 97.21% for high-risk scenarios, which is much higher than the traditional convolutional neural network's 86.04% and86.57%, respectively. In practical applications, the manual review rate of the AI animation sensitive content recognition system is only 5.27%, with an update cycle of 3.5 days. It can be seen that the system has a high accuracy in identifying sensitive content and a low cost of manual review. The method proposed by the research institute provides a high-precision and low-cost solution for the security review of AI animation content, which can effectively ensure the health and safety of the network environment.