Page 43 - CCFA Journal - Seventh Issue

加中金融 机器学习 Machine Learning

Traditional fraud detection or AML monitoring systems utilize business rules meticulously designed by subject-matter experts or rely
on manual reviews conducted by trained analysts. Due to extensive human intervention required, such systems are difficult to
maintain or expand. They are also less responsive to emerging trends in criminal behaviors. A criminal may even circumvent strictly
rule-based detection if rule details are leaked. In contrast, ML can learn from data without being explicitly programmed. It is capable
of recognizing complex patterns not apparent to human eyes and is thus more sensitive to new criminal activity patterns. As
computing power is now cheap and readily available, after proper hardware and software deployment, ML-based systems can deliver
scalable performance. Furthermore, an expanded throughput brings more data, which can be used to continuously improve ML
algorithms to attain higher model precision. The automation and fast responses enabled by ML-based systems allow FIs to enhance
their efficiency and reduce operation costs.

In practice, ML models are often used as one component of the entire fraud or AML monitoring process instead of standalone decision
tools. Model output (typically risk scores) is fed into downstream business rules or procedures to determine how certain events
should be treated. For example, it may help decide in real time whether a user’s online payment should be declined. In another
scenario, it may alert an analyst that a fund transfer deserves investigation for money laundering. Properly developed ML models can
automate decision making and significantly reduce the burden on analysts so that they can focus on reviewing riskier activities. FIs’
fraud detection and AML monitoring exercises complemented by ML models are more flexible, efficient, and robust.

ML models pose new challenges to model risk management at FIs. On one hand, regulators expect FIs to clearly understand their
models, but on the other hand, some ML algorithms are rather complex. The results from a model based on neural networks, for
instance, are much harder to interpret than those from decision trees. In the case of vendor models, the vendors tend to share only
limited high-level model information with FIs. To demonstrate ML models’ effectiveness with such incomplete information requires
creativity. Model risk management frameworks at FIs must catch up with the rapid development of ML models.

The playing field of fraud and AML analytics is expanding. It is projected that the global fraud detection and prevention market will
grow from $27 billion in 2021 to $75 billion by 2028 [4]. Meanwhile, the global AML market size is also forecasted to increase from
$2 billion in 2020 to $5 billion by 2027 [5]. Multiple studies suggest that fraud detection and prevention are among the most common
reasons for businesses’ adoption of AI [6]. A recent report by SAS and KPMG shows 57% of the institutions surveyed have already
deployed AI or ML in AML production or have plans to do so in the near future [7]. Given wider adoption of AI among FIs, one can
expect ML to be a key driver behind this growth in fraud and AML analytics.

While extensive ML research and development constantly offer innovative tools to FIs, criminals can also adopt them in their illegal
activities. In a 2020 swindle, fraudsters used deep voice technology to clone the voice of a company’s director and easily tricked a
bank into transferring $35 million [8]. In the foreseeable future, the battle against financial crimes will unavoidably involve the struggle
between AI technologies wielded by both the good and the evil.

传统的欺诈检测或反洗钱监控系统主要利用由专家精心设计的业务规则，或依赖经训练的分析师进行的手动审查。由于需
要大量的人工干预，这样的系统难以维护或扩展， 也无法及时发现犯罪行为的新兴趋势。如果规则细节被泄露，犯罪分
子甚至可以规避严格基于规则的检测方法。相比之下，机器学习无需明确编程即可从数据中学习，它能够识别人眼看不到
的复杂模式，因此对新的犯罪活动模式更加敏感。由于计算能力现在便宜且容易获得，在适当的硬件和软件部署之后基于
机器学习的系统可以提供可扩展的性能。此外，扩大的吞吐量可以带来更多的数据，这些数据可以用于不断改进机器学习
算法以获得更高的模型精准度。基于机器学习的系统实现的自动化和快速响应使金融机构能够提高效率并降低运营成本。

在实践中，机器学习模型通常被用作整个欺诈或反洗钱监控过程的一个组成部分，而不是独立的决策工具。模型输出（通
常是风险评分）被输入到下游业务规则或程序中，以确定应如何处理某些事件，例如它可以帮助实时决定是否应该拒绝用
户的在线支付。在另一种情况下，它可能会提醒分析师某一笔资金转移值得进行洗钱调查。适当开发的机器学习模型可以
自动化决策并显著减轻分析师的负担，以便他们可以专注于审查风险较高的活动。金融机构的欺诈检测和反洗钱监测活动
辅以机器学习模型能变得更加灵活、高效和稳健。

机器学习模型对金融机构的模型风险管理也带来了新的挑战。一方面，监管机构希望金融机构清楚地理解他们的模型，但
另一方面，一些机器学习算法相当复杂难以理解。例如，基于神经网络的模型的结果比决策树的结果更难解释。在供应商
模型的情况中，供应商倾向于仅与金融机构共享有限的大致的模型信息，用这种不完整的信息难以证明机器学习模型的有
效性。金融机构的模型风险管理框架必须跟上机器学习模型的快速发展。

欺诈和反洗钱分析的应用正在扩大。全球欺诈检测和预防的市场预计将从 2021 年的 270 亿美元增长到 2028 年的 750 亿美
元 [4]。同时，全球反洗钱市场规模也预计将从 2020 年的 20 亿美元增加到 2027 年的 50 亿美元 [5]。多项研究表明，欺诈
检测和预防是企业使用人工智能的最常见原因之一 [6]。 SAS 和毕马威最近的一份报告显示，57% 的受访机构已经在反洗
钱生产中使用了人工智能或机器学习，或者计划在不久的将来应用 [7]。鉴于人工智能在金融机构中的广泛采用，机器学
习预计将成为欺诈与反洗钱分析增长背后的关键驱动力。

虽然广泛的机器学习研究和开发不断为金融机构提供创新工具，但犯罪分子也在其非法活动中改进采用新技术。在 2020
年的一起骗局中，欺诈者使用深度语音技术来克隆公司董事的声音，并轻松欺骗银行转移 3500 万美元 [8]。在可预见的未
来，打击金融犯罪的斗争将不可避免地涉及善与恶的人工智能技术之间的斗争。

CCFA JOURNAL OF FINANCE February 2022
Page 43 第43页