Fraudsters are known to actively collaborate; sharing both data and knowledge to increase their chances of success, or simply to amplify their attack vectors. Why do financial institutions not act similarly to counter these threats?
Data sharing between financial institutions can certainly help them address business challenges, including fraud prevention. But this has always been strictly restricted due to competition implications, client confidentiality concerns and data privacy regulations. This is especially true where such sharing involves customers’ personal data. On top there may also be reputational risks or concerns. As a result, collaboration between financial services firms for fraud prevention purposes has historically been low. Even when it has happened, it has typically only been on an ad hoc, bilateral and reactive basis. To overcome these challenges, we are applying an innovative approach we term Collective AI.
This is based upon our belief at NetGuardians that certain aspects of fraud prevention are fine candidates for safe, secure, privacy-respecting consortium models and techniques. We believe that it can bring significantly improved outcomes for participants.
We compute and generate anonymous signals and statistics from certain of the AI models that our subscribers operate, as well as the users’ interactions with these model outputs (e.g. ‘fraud, ‘not fraud’ classifications). These signals and statistics can then be automatically shared across subscribers – on a fully secure basis – for the other subscribers to incorporate automatically in their equivalent models. All within the network of participants benefit from others. They can learn from the actions of their peers. These consortium models operate without any violation of data protection and legal frameworks.
How does it work? An example may help illustrate
In the interconnected world of financial services, many banks and corporates are sending money to the same counterparties or beneficiaries. These firms are quite likely to be analyzing and profiling their counterparties, their beneficiaries and even the underlying accounts at these beneficiaries. But this is in a segregated fashion; assigning a risk score based on the knowledge they have at their disposal, such as the frequency of payments to these institutions, their accounts, the value distribution, and common currencies.
This may create a pattern of known payment behaviour to this counterparty or beneficiary account such that they can be considered ‘trusted’, or a low fraud risk to them.
But another firm making a payment to the very same beneficiary or account may have no such history to reference. It may be their first time making a payment to this account. It is valuable insight to understand if their peers have previously paid to this account safely.
All participants in the consortium can leverage the collective intelligence of their peers, across a far broader data context; safely, securely and anonymously. Model performance is significantly improved; false positives reduced. Specific consortium level statistics are anonymously generated and injected into models as an additional feed of features in a segregated system.
This approach has the power to transform the performance of certain fraud prevention models, by bringing the following innovations:
Federated analysis technique allows the participants in the consortium to share the insights from their analysis without sharing the data itself. This technique makes our approach compliant with the highest security data standards and avoids any violation of data protection laws.
Participants in these Collective AI models are banks, corporates, fintechs and neobanks. With different business focus (e.g retail, digital, wholesale, private banking) and diverse regional coverage. This heterogeneity this provides is a key element for continuous improvements in fraud prevention rates.
Meta-data signals are generated automatically (without need of human intervention or specific reporting) and incorporate the full knowledge of the consortium on multiple different dimensions. These meta-data signals are dynamically computed on an ongoing basis. This is very far from the binary ‘yes/no’ of block lists and accept lists.