Finding Signature Matches in Corporate Documents

PUBLISHED ON

Associating signatures with companies

One of the many tools in the arsenal of a money laundering investigator is the ability to check signatures across company documents and find signature matches as a way to identify AML case risk levels. Individuals required to sign documents for authentication purposes (such as accounts records) leave traces of their identity across the companies they associate with and this can be a valuable source of insight. In some cases, the count of their signature alone can raise suspicion, as in the case of Ali Moulaye, whose signature decorates thousands of UK companies. But the association of companies to a particular signature can have many other uses and offer a wealth of information around the signatory.

High volume of companies involved in money laundering

Whereas certain individuals must be identified clearly as officers on registers such as Companies House, many signatories of company documents are not identified in a computer-readable way, and this significantly raises the time costs of tracing a particular individual’s association across various companies. Where a particular signature is of interest, whether or not it is signed authentically, AML investigators must choose whether to invest hours of their time obtaining documents, scanning for signatures, and identifying positive matches with their sought-after signature. Since money laundering activity traditionally uses large numbers of companies, and therefore documents, to obfuscate its evidence trail, this can easily reduce the task of tracing an individual to a lottery of choosing a rewarding sample of documents, or take it out of reach completely.

Automated signature matches

Using a pipeline comprising a combination of heuristic methods and machine learning techniques can bring this full task back into the reach of investigators. Such a pipeline automatically retrieves documents, extracts signatures and partitions them by author at a significantly faster rate than a human investigator, raising the investigator’s ability to scale up at the same rate as the money launderer’s ability to generate the trail.

While signature investigations all involve the same core subject, there are different types of investigation that might be useful, depending on the level of prerequisite information and the openness of the question. For example, if we have a number of signatures of interest, and a set of companies likely to contain them, then a pipeline run will check for matches (signatures that are clustered together) with the known signatures, and return the complete clustering of all others. Alternatively, we might just start with a set of company names, or documents, and inspect the clustered signatures for any anomalous results: a high number of signatures for one signatory, multiple occurrences of the same signature signing for different identities. The pipeline dramatically reduces the time spent performing mundane tasks so investigators can skip to the results.

Augmented investigations

It is important to note that this is not a replacement for the vital work of human investigators, but rather an aid. Human speed is surpassed, but human accuracy is still unparalleled. And it remains that with an increase in the scale of documents analysed, maintaining accuracy of results, human or machine, is increasingly difficult. Experts can obtain high accuracy in identifying a small number of distinct signatures, but if the task were to identify a signature as one of thousands, confidence would be in much shorter supply.

Furthermore, signatures in company documents are not made with the intention of being identifiable, only verifiable. They are often obscured by stamps, text, and other obstructions. They vary significantly not only by person but by time – two signatures from different people can easily look more similar than two signatures by the same person separated by a month. These problems are non-trivial even for human investigators, and pipeline resorts to automatic signature cleaning to tackle them.

Much research has focused on the problem of signature verification, where one signature must be compared against exactly one known signature to decide whether they share a signatory – a one-to-one comparison. Yet the problem of signature identification, a many-to-many problem, is arguably much more promising to the field of anti-money laundering investigation, and there is undoubtedly further progress to be made.

Read the full research paper here: Fully-Automatic Pipeline for Document Signature Analysis to Detect Money Laundering Activities