Tax fraud, social fraud, credit card fraud, ... the types of fraud are varied and with the diversification of the means of payment, they multiply and become more complex, harming individuals, companies and the State for exorbitant costs.
Thus, in 2020, payment fraud amounted to 1.28 billion euros* (and 644 million euros for the first quarter of 2021 alone) and estimates of tax fraud vary between 80 and 100 billion euros**. The fight against fraud has become a major challenge.
To counter this scourge, the State, companies and banks are seeking to strengthen their anti-fraud measures and are increasingly turning to the use of Artificial Intelligence. The Ministry of Economy and Finance claims to have increased its performance in detecting tax fraud through the use of datamining algorithms. AI is undoubtedly a key ally in strengthening the fight against fraud, provided it is used properly!
The fight against fraud, a good case study to understand the potential of AI. At the heart of the system: the business expert.
Upstream, analyze the context well, for an optimal data modeling.
It should be noted that fraud, although extremely costly, remains a rare event. And the modeling of rare events requires a very high standard of data preparation, and to take into account all the weak signals.
As with any modeling project, the first step is to create a relevant dataset but also to understand the context of the fraud: What is the purpose of the fraud? How does the fraudulent tool work? Where are the risks of compromise? Who can commit fraud? How can the fraudster recover his gain?
A meticulous understanding of the context in which the company evolves, based on a detailed knowledge of the business, will make it possible to know which types of fraud are targeted and to create algorithms that meet the operational constraints of the frauds to be detected.
For example, in the fight against credit card fraud, it is necessary to know that fraud reports are not immediate and that the kinematics of fraud (operating modes) are constantly evolving. These elements must be taken into account when designing the algorithm and modeling. Otherwise, there will be biases between the POC phase and the production of the algorithm, which will lead to results that do not meet expectations.
Once all the context and constraints are fully understood, how can we best use the power of AI?
AI has many contributions to make in the fight against fraud, especially when coupled with human expertise. AI opens up new possibilities for analysis and performance for the expert by allowing him to :
- Optimize existing anti-fraud mechanisms:
Before using machine learning algorithms, many organizations have already implemented anti-fraud systems, often based on rules created by experts according to their observations and field feedback. AI can challenge these rules and propose optimizations (e.g. new thresholds) in order to gain relevance and reduce false positives that correspond to lost earnings (e.g. wasted investigation time, wrongly blocked financial transactions...).
- Identify new fraud kinematics
AI can detect new fraud patterns. To do so, it is essential that the results of the algorithms used are interpretable by an expert in order to understand the new fraud kinematics identified/generated by the algorithm. Indeed, the so-called "black box" models (e.g. xgboost) may be relevant for detection but are not relevant for explaining and understanding the kinematics of fraud. The more complete the understanding of the fraud kinematics, enriched by human experience and knowledge, the more relevant and sustainable the response to this fraud will be.
- Gain in reactivity and vigilance
Fraudsters are quick to adapt: as soon as they notice that a fraud pattern has been contained by new controls, they do not lack imagination to create another one. So you have to be vigilant and reactive! AI allows to quickly detect new uses or anomalies in behaviors (e.g.: appearance of a new e-commerce site or strong increase of activity on a site) and to report them. Be careful, an anomaly does not necessarily correspond to a fraud (e.g.: it can only be an increase in activity on a site following the release of a new product or a promotional campaign), so you should always submit the results to a fraud expert.
In several aspects, AI is a lever to optimize fraud detection, but like any intelligence, it needs to be "educated" to the context of fraud that we want to fight, to be relevant and express its full potential. It is therefore essential to combine the power of AI with business expertise to have the most complete, responsive and efficient anti-fraud system possible.