The GANfather: Controllable generation of malicious activity to expose detection weaknesses and improve defence systems.

Criminal activities are typically adversarial in nature, where an attacker and a defence system are constantly adapting to each other's behaviour. If the defence systems are helped by automated detection methods, then those methods need to be updated frequently. In practice, this means that the defence systems are always one step behind the attackers. For example, in anti-money laundering systems, new labels representing suspicious activity are frequently delayed by weeks or months and some money laundering activity may never be found, leading to detection systems that are inaccurate and resulting in an estimated undetected €0.7-3 trillion being laundered annually. To tackle the problem of missing or delayed labels in adversarial settings, we propose The GANfather, an adversarial and label-free method to both (1) generate a variety of meaningful attacks, as guided by a custom, user-defined objective function; and (2) train a defence system to detect such attacks. Optionally, we can ensure that the generated attacks escape an existing detection system, revealing current weaknesses which the new defence system actively corrects. Our method is inspired by generative adversarial networks (GANs), but unlike GANs we nudge our generator to produce out-of-distribution data using a loss function that characterises criminal activity. Importantly, our method does not require any labelled examples. We test our framework in two real-world use-cases, namely injection attacks in recommendation systems and anti-money laundering. In the former, we show how an injection attack with a limited number of generated fake profiles is sufficient to successfully recommend an item to a large number of users. These generated injection attacks are more effective in recommending the target item than naive ‘bombing’ strategies and harder to detect. In the latter, the generated attacks are able to simulate money laundering and move cumulative amounts close to 250 thousand dollars through a network of accounts without being detected by existing systems. We also show how we can train a new defence system that captures all these synthetic attacks, potentially saving millions of dollars in detected criminal activity. Our method is generic and applicable in a variety of adversarial domains, exposing current liabilities with the generated data and strengthening the defence systems against current and future malicious attacks.