Partners in Crime: Consilient and Intel Link to Improve Automatic Fraud Detection Mechanisms, Lowering the Rate of False Positives Currently Hampering the Industry

In this most connected of worlds—with whole buildings allocated as data centers where smoke is constantly coming out, belying the fact that there can be dozens of air-conditioning units inside these buildings to keep the computers cool—thousands of financial transactions make their way through international payment systems every second of every day. This naturally brings in all sorts of benefits for the world’s economy, but it also attracts a fair amount of financial crime and practices that are clearly detrimental to societies—including, for example, the funding AND money laundering of criminal enterprises and terrorism. Enter Consilient and its partnership with Intel. They have created an approach to combat financial fraud—one encompassing federated learning and application isolation technology, the latter meant to put up barriers around sensitive processes and models—with the aim of improving the current and quite clearly unacceptable 95% rate of false positives that fraud detection processes, well, boast. Formed in 2020 by the joint efforts of K2 Integrity and Giant Oak, Consilient is focused on establishing an international collaborative effort where insights generated by the just-described setup can be widely shared among financial institutions, thereby detecting patterns of money laundering more readily and in so doing vastly reducing the current rate of false positives.       

According to the United Nations, around 2% to 5% of the world’s annual gross domestic output is laundered, meaning that vast amounts of money obtained from illicit activities such as drug trafficking and corruption—between US$800 billion to US$2 trillion, in fact—are being funneled into the economy every year. This is certainly a significant problem, and considering the size and speed of financial transactions, a very difficult one to tackle—direct human intervention cannot be expected to follow, let alone discern, any patterns in such a gigantic ocean of information. While this is a problem perhaps ripe for pattern detection systems—an issue that Machine Learning (ML) tends to excel at, considering that ML is at heart an exercise in “curve fitting”: namely, the construction of a model that best fits a series of data points—the results are currently rather poor. The rate of false positives is far too high to be of any use; this is where Consilient and Intel enter the fray.

The Consilient model platform—powered by Intel’s Software Guard Extensions (SGX), itself built into Intel Xeon Scalable processors—carries out a federated kind of ML where each ML model can be trained at a given organization using local data. The model is then shared with other organizations where further training can be conducted. In each case, the model is being trained on local data that do not travel with the model itself (i.e., the data are not shared). This is meant to protect privacy, and this works well with the application isolation technology underlying the Intel SGX engine because the latter works by protecting selected code and data from any external modification. The Intel SGX engine partitions applications into hardened “enclaves” or modules such that there are no centralized data at any point. In this case, the overall setup works by scoring customers and searching for both normal and abnormal patterns in the data sets. The relevant models are trained on different data sets as the so-called master model is reiterated. This is where insights are eventually generated. The hope is that the insights will be effectively shared across financial institutions—insights that are based on the most up-to-date risk models, given that the algorithms are traveling as it were from one data set to another, collecting relevant information along the way to identify irregular patterns. The result, according to Consilient, is a significant reduction in false positive rates (a huge selling point) compared with locally run ML models that are the norm today and that in some tests score below the 60% mark.

Financial crime comes in all shapes and forms. Money laundering is just one type; there are various types of fraud (e.g., bank fraud, insurance fraud, medical fraud, etc.), theft, scams, tax evasion, bribery, embezzlement, forgery, and so on. As such, there is a large number of scenarios when it comes to the illicit financial transactions that take place in the world on a daily basis. This necessitates a multipronged approach to tackling the overall phenomenon—meaning assistance from legislation (e.g., the U.S. Sarbanes-Oxley Act of 2002, which expands reporting requirements for all U.S. public companies) and law enforcement agencies, and in particular revenue agencies. In addition, banks and card companies are major players in this fight against financial fraud and are certainly invested in stopping fraudulent payments; in 2020 alone, U.K. banks managed to stop close to US$2 billion worth of unauthorized fraud losses through well-designed protocols.

Consilient’s method, focused on scoring bank customers and detecting abnormal patterns from this, is but a small cog in the fight against financial crime. While its federated learning models are seemingly capable of improving the incredibly high rate of false positives of current methods, it must be realized that financial fraud may not just be a technical problem for ML to solve via a federated method or not. This setup may help prevent the blocking of legitimate transactions more effectively than it may block fraudulent ones per se, and that is a significant achievement for the technology. The Intel SGX engine, in particular, is an apt tool for keeping sensitive algorithms and data secure and private at every level. In this vein, ABI Research recommends that financial institutions participate in model training across the industry and share the generated insights.