“类人”社交机器人检测数据集扩充方法研究

该文提出了基于海林格距离的变分进化生成式对抗网络(HVE-GAN)，实现“类人”社交机器人检测数据集的扩充。HVE-GAN将进化生成式对抗网络(E-GAN)的生成器修改为变分自编码器(VAE)结构，提高了生成数据的“真实性”及多样性程度；将E-GAN生成器Heuristic损失函数更改为改进的海林格距离，在训练过程中加快了模型收敛速度、稳定了生成器的梯度，避免了不稳定的训练过程影响生成数据质量。实验结果表明，利用HVE-GAN模型生成的“类人”社交机器人数据的“真实性”与多样性程度均明显优于基线模型。

Research on Expansion Method of Detection Dataset for"Human-like" Socialbots

A Hellinger distance based variational evolutionary generative adversarial networks (HVE-GAN) is proposed to expand the detection dataset of “human-like” Socialbots. HVE-GAN modifies the generator of evolutionary generative adversarial networks (E-GAN) to a variational autoencoder (VAE) structure to improve the “authenticity” and diversity of the generated data, and changes the Heuristic loss function of the E-GAN generator to an improved Hellinger distance to speed up the model convergence during the training process, stabilize the gradient of the generator, and further avoid unstable training processes that affect the quality of the generated data. Comparative experimental results show that the “authenticity” and diversity of the “human-like” social robot data generated by the HVE-GAN model proposed in this paper are significantly better than the baseline models.