Perverted Justice Dataset

The risks to children of online predators in real time gaming environments have been an area of growing concern. Research towards the development of near real time capabilities has been the focus of most queries published in this area of study. In this paper, we present Protectbot, a comprehensive safety framework used to interact with users in online gaming chat rooms. Protectbot employs a variant of the GPT-2 model known as DialoGPT, a generative pre-trained transformer designed specifically for conversation. By generating content that closely resembles human dialogue, DialoGPT allows Protectbot to engage users in interactive chat sessions. At the end of each chat, Protectbot analyzes the user's messages to identify any indications of potentially predatory behavior, enhancing the platform's capacity to safeguard its users. Protectbot architecture implements a text classifier that was trained and tested on the PAN12 dataset for identifying sexual predators. fastText word embeddings are generated from the chat text and aggregated into sentence vectors, which are then used as input features to train an SVM classifier. The proposed model achieved notable performance metrics, with a recall, accuracy, F1-score, and F_0.5-score of 0.99, marking a significant improvement over previous methodologies. A new dataset is prepared based on 71 predatory chats obtained from Perverted Justice (PJ), to evaluate the classifier's performance. The proposed approach demonstrates a high true positive rate of classifying predatory behavior by replacing the SVM with the KNN classifier 

实时游戏环境中，在线性诱骗者对儿童构成的风险已成为日益受到关注的研究议题。该研究领域已发表的绝大多数相关研究，均围绕开发近乎实时的防护能力展开。本文提出Protectbot——一款面向在线游戏聊天室用户交互的综合安全防护框架。Protectbot采用了GPT-2的一款变体模型DialoGPT，这是一款专为对话场景设计的生成式预训练Transformer模型。通过生成与人类对话高度相仿的内容，DialoGPT可使Protectbot与用户开展交互式聊天会话。在每场聊天结束后，Protectbot会对用户的聊天消息进行分析，以识别潜在的诱骗行为迹象，进而提升平台守护用户安全的能力。Protectbot的架构集成了一款文本分类器，该分类器基于PAN12数据集进行训练与测试，用于识别性诱骗者。研究人员从聊天文本中生成fastText词嵌入，并将其聚合为句子向量，随后将这些向量作为输入特征训练支持向量机（SVM）分类器。所提出的模型取得了优异的性能指标，其召回率、准确率、F1值以及F_0.5值均为0.99，较此前的同类方法实现了显著提升。研究团队基于从Perverted Justice（PJ）获取的71条诱骗式聊天记录构建了全新数据集，用于评估该分类器的性能。通过将SVM分类器替换为K近邻（KNN）分类器，所提方法实现了更高的诱骗行为识别真阳性率。