Machine Learning at Veterans Benefits Administration

Consumer data breaches and cyber threats have become a growing concern in recent years, prompting federal agencies to act to prevent identity theft and identity-based criminal activity. The Veterans Benefits Administration (VBA) partnered with Booz Allen to stand up a machine learning-based fraud detection system to identify suspicious changes made to federal beneficiaries’ financial information. As a result, $2.4 million was protected in fiscal year 2019.

VBA distributes federal benefits to qualifying beneficiaries of the Department of Veterans Affairs (VA)—often veterans and their family members. In cases of identity theft or fraud—if a veteran reports a stolen ID or monthly benefits check, for example—the agency has historically implemented a “pay and chase” model.

Moreover, the agency has struggled with the downstream implications of fraud and identity theft. While the anti-fraud program has been successful in deterring criminal activity, high-profile fraud attempts, such as those against Purple Heart recipients or high-level government officials, or those made by internal VA members (e.g., employees), can create widespread media coverage and huge political backlash.

VBA knew there had to be a better way to prevent fraud. How could the agency verify a veteran’s identity before letting them change financial information? How exhaustive should the identity authentication process be? Is 2-factor authentication secure enough, or is biometric data needed? 

To protect the veteran population against criminal activity and save taxpayer money, VBA knew it needed to find a balance—ensuring the agency could protect the veteran population from fraud, while making sure their protection measures remain relatively non-prohibitive for veterans.

Our team of data scientists and strategy consultants helped VBA implement a more comprehensive deterrence strategy. It allows the agency to physically prevent stolen dollars from impacting VA ledgers and compress criminal threat paths by using transactional data captured in internal databases to find criminal behavior in a statistical manner. 

Specifically, the system aims to identify fraudulent changes made to a beneficiary’s bank account information, allowing a criminal to route a veteran’s monthly benefits to a fraudulent bank account. Identity theft and fraud detection is, by nature, an adversarial problem, so machine-learning tools serve as good solutions.

The VBA system can learn fraudsters’ behavioral tendencies as they evolve over time and continuously flags risky transactions for remediation. This preventative capability has helped VBA rethink its anti-fraud strategy and institute modern frameworks to solve complex problems, protect the veteran population, and save taxpayer dollars.