AI based Anti-fraud systems – advantages and threats

Challenges facing modern anti-fraud class systems

The world with prevailing access to the Internet, mobile electronic devices is not only the domain of entertainment and business.  Financial market knows  it  and following  the requirements of the modern world.

The impact of mobile technologies on the shape of the market offering financial services is very significant.
Banks and other financial institutions are increasingly turning away from traditional customer service methods for mobile technologies.
Nowadays, access to a bank account or other types of financial services, through applications on mobile phones or websites is a the standard and must-have for almost every customer.

These modern channels of communication with a financial institution apart from obvious benefits for the client and the institution itself also carry a huge risk of using them for criminal purposes.
The bank robbery known from the world of the film, with a gun in hand, is slowly becoming a thing of the past.

The real threat to  a a  bank today   is a hacker equipped with a laptop and Internet access that prepares phishing campaigns for clients of the attacked bank or uses social engineering to obtain certain information from the client, for example by  a regular phone call claiming to be a bank employee.

With the increasing importance of anti-fraud systems, unfortunately, the awareness of criminals and the range of techniques used by them is also growing.


A compromise between security and the convenience of using tools in the context of anti-fraud.

One of the basic challenges facing modern systems preventing the use of financial systems for criminal activity, including money laundering, terrorism financing, and unauthorized operations  (fraud) is the need to ensure operation of those system in real time or near real time.

Most of the currently offered systems of this type base its activity on selecting potentially fraudulent transactions and then suspending their action (decision) until the manual approval of such a transaction or rejection by an employee of a financial institution or bank.

It is obvious that this approach creates some discomfort for the users of the financial system as it can often lead to unjustified blocking of funds or suspension of the correct transfer of funds, sale of a financial instrument, etc.
All financial institutions, including banks, cantors, exchange brokers, payment intermediaries, factoring companies, and insurance companies face the challenge of finding a compromise between the convenience of using the system and its security and the security of the financial institution itself.

This problem is particularly important in the context of stock exchange brokers acting in financial operations on regulated markets such as stock trading, futures contracts, currency options and bonds.
High-frequency trade (HFT) is perhaps the biggest technological challenge in the field of optimization and effectiveness of anti-fraud systems that we are currently dealing with.


Anti-fraud analytics in real time

In the context of the current financial markets in HFT trade in which we deal with operations performed within milliseconds and millions of operations per day, the obvious conclusion is that anti-fraud systems can not be based on semi-automatic approval by a human (institution’s employee).

The need to eliminate the human factor from this process seems obvious, so the question arises whether the automation of the process is possible to the extent that the anti-fraud system was not mistaken in both false-positive and false-negative.

Modern algorithms based on machine learning including Deep learning, which in the popular literature are called Artificial Intelligence (AI) are an important step forward towards full and efficient automation of anti-fraud processes due to its unmatched ability to adapt and learn from test collections.

An unquestionable advantage of these systems is that the analyst or system designer responsible for security does not have to declare or design potential paths with truth, but only provides a set of operations performed by users over the years and each operation before its approval assesses in terms of probability value whether it is an operation correct whether there is a criminal operation.


The performance of AML and Anti-fraud systems based on machine learning.

The second major advantage of systems based on machine learning is the speed of their operation compared to classical algorithms.
It is true that the learning process itself is computationally demanding but the production application of AI when we deal with the well-trained CNN or LSTM neural network model allows us to get response (decision) in milliseconds after request.

Thanks to the use of machine-based algorithms, we can also constantly expand the scope of data that is included in the analysis of potential factors that may indicate that we are dealing with fraud.

In addition to parameters such as the IP number, location, transaction amount and frequency of activities – we can take into account a larger number of variables.

A properly designed and taught neural network is able to analyze in parallel dozens of different input parameters, including biometric data and behavioral data of users, and compare them with the saved pattern.

Examples of such data may be:

  •  a unique pattern for using input devices, for example a mouse keyboard
  • user-unique way of authorization and performance After logging in to a platform offering financial services
  • And others that depend on the data set we have access to in the context of the system user.

Threats resulting from the use of artificial intelligence in transaction analysis processes.

The development of artificial intelligence, including machine-based algorithms, will probably affect the quality of anti-fraud systems and will allow to increase the scope of data to be analyzed and for more accurate profiling of the user for later identification.

In the context of threats arising from the use of machine-based systems to make fully automated decisions, in the context of Anti-fraud and AML, it should be mentioned that the criminal behavior scheme will also be adapted to such systems.

Scheme of attack on a self-learning anti-fraud class system (including AML)

The awareness and technical skills of criminals grows just as their  knowledge of how artificial intelligence works.

It is certain that criminals will try to take actions that deliberately distort the behavioral model. The purpose of these activities will be to reduce the effectiveness of the model itself.

Such goals can be achieved by clever social engineering or a network of accounts controlled by the attacker (botnet), which will carry out operations at the limit of the likelihood of fraud. With a sufficiently large amount of such activities – the self-learning model will start to recognize them as a positive and safe action.

Subsequently, the scheme of such an attack may consist in carrying out operations not much different from dozens of previous operations, which were not classified as fraud, and finally the actual fraud operation will be passed through the decision filter.

This is of course only a certain scenario, assuming that the network will learn itself without supervision. Of course, in the case of an advanced system, the supervision of a properly educated expert on the process of feedback learning of the model must be carried out continuously.
The purpose of such control is to eliminate artificially generated behaviors that would give the impression of universal actions.