An Interview with Massieh Najafi: The Challenges of Automating Analytical Detection

Alexa Rzasa
by Alexa Rzasa
category Perspective

We recently had the opportunity to chat with Massieh Najafi, executive VP and head of the data science department at Guardian Analytics. He is a pioneer in applying behavioral analytics to fraud detection and security problems. He is the main architect of Guardian Analytics core technology, employing machine learning and big data technologies to develop scalable behavioral analytics-based fraud detection solutions.

Under his leadership, Guardian Analytics has become a frontrunner in artificial intelligence (AI) based fraud detection technologies for banking systems and payment industries, serving more than 450 banks and financial institutions. He has a broad experience and interest in security, payment, banking systems, FinTech, and Blockchain, etc. Massy holds a Ph.D. from the University of California, Berkeley.

We wanted to get Massy’s expert perspective on the major challenges in automating analytical detection, symmetries between fraud and cybersecurity and more.

As a highly accomplished data scientist, which part of the problem do you find most interesting, the data or the science, and can you explain why?

I think both sides are interesting and equally important. On the data side, it requires skill and passion to digest and extract insight from data. You need to enjoy spending time with data and be patient to get data to talk to you. On the other hand, the science side is a combination of art and technical knowledge. It takes skill and creativity to model and project data based on its characteristics, your objections, and various limitations.

As an expert in fraud detection, what do you think are the most interesting symmetries between fraud and cybersecurity?

The main symmetry between fraud detection and cybersecurity is the adversarial aspect of it. Unlike other classification problems, in security and fraud detection you are dealing with adversarial systems (e.g., bad guys, fraudsters) that know about your mission (to detect and stop them) and will do their best to find a way to game or fool your models/algorithms to stay under the radar while achieving their goals.
This makes the problem more challenging, as it is no longer sufficient to design based on the current status quo. You also need to consider the bad guys’ new strategies in response to your design/solution. In other words, a fraud detection or security problem is the battlefield of two intelligent players: the bad guys (who have the intention of doing harm or stealing something) and protectors whose aim is to design solutions/algorithms to detect and stop bad guys.

What do you think are the major challenges in automating analytical detection?

An automated analytical detection solution needs to consider different parameters to assess the fidelity of a new observation. As the number of parameters increases, the Curse of Dimensionality issue arises, which means the amount of data needed to train/build the model explodes exponentially. The Curse of Dimensionality issue is more challenging in security and fraud detection problems, as by their nature, in such problems we are already dealing with extremely imbalanced data.

What interested you in becoming a strategic advisor to Respond Software?

I have a general interest in helping AI-based companies with a fresh look at long-standing problems. Respond Software has approached the enterprise security problem in a different way. Rather than formulating it as a conventional classification problem, they have defined it as a likelihood estimation problem. This strategy gives them a lot of flexibility to deal with challenges such as dimensionality, imbalanced data, adaptiveness, and customization. I believe their technology has the potential to introduce new horizons into enterprise security and put protectors ahead of the game with bad guys.