In recent years, with the rapid development of digital currency, digital currency brings us convenience and wealth, but also breeds some illegal and criminal behaviors. Different from traditional currencies, digital currency provides concealment to criminals while also exposing their behavior. The analysis of their behavior can be used to detect whether the current digital currency transaction is legal. There is a problem that most digital currency transactions are in compliance with laws and regulations, and only a small part of them uses digital currency to conduct illegal activities. It belongs to the problem of sample imbalance. It is quite challenging to accurately distinguish which transactions are legal and which are illegal in the massive digital currency transactions. For this reason, this study combines the mutual information and the traditional cross-entropy loss function and obtains the loss function based on the mutual information prior. The loss function based on the mutual information prior is that the bias of the category prior distribution is added after the output of the model (before the softmax), which makes the model consider category prior information to a certain extent when predicting. The experimental results show that the use of the loss function based on mutual information prior to the detection of digital currency illegal behavior has a good effect in SVM, DNN, GCN, and GAT methods.

中文翻译：

---

基于互信息先验损失的数字货币非法行为检测

近年来，随着数字货币的飞速发展，数字货币在给我们带来便利和财富的同时，也滋生了一些违法犯罪行为。与传统货币不同，数字货币为犯罪分子提供了隐蔽性，同时也暴露了他们的行为。对其行为的分析可以用来检测当前的数字货币交易是否合法。存在的问题是，大多数数字货币交易符合法律法规，只有一小部分使用数字货币进行非法活动。属于样本不平衡问题。在海量的数字货币交易中，准确区分哪些交易是合法的，哪些是非法的，是非常具有挑战性的。为此原因，本研究将互信息与传统的交叉熵损失函数相结合，得到基于互信息先验的损失函数。基于互信息先验的损失函数是在模型输出之后（在softmax之前）加入类别先验分布的偏差，这使得模型在预测时一定程度上考虑了类别先验信息。实验结果表明，在检测数字货币非法行为之前使用基于互信息的损失函数在SVM、DNN、GCN、GAT方法中都有很好的效果。基于互信息先验的损失函数是在模型输出之后（在softmax之前）加入类别先验分布的偏差，这使得模型在预测时一定程度上考虑了类别先验信息。实验结果表明，在检测数字货币非法行为之前使用基于互信息的损失函数在SVM、DNN、GCN、GAT方法中都有很好的效果。基于互信息先验的损失函数是在模型输出之后（在softmax之前）加入类别先验分布的偏差，这使得模型在预测时一定程度上考虑了类别先验信息。实验结果表明，在检测数字货币非法行为之前使用基于互信息的损失函数在SVM、DNN、GCN、GAT方法中都有很好的效果。