Researchers Create New Approach to Detect Brand Impersonation

Security researchers have designed a new way to detect brand impersonation using Siamese Neural Networks, which can learn and make predictions based on smaller amounts of data.

These attacks, in which adversaries craft content to mimic known brands and trick victims into sharing information, have grown harder to detect as technology and techniques improve, says Justin Grana, applied researcher at Microsoft. While business-related applications are most often spoofed in these types of attacks, criminals can forge brand logos for any organization.

"Brand impersonation has increased in its fidelity, in the sense that, at least from a visual [perspective], something that is malicious brand impersonation can look identical to the actual, legitimate content," Grana explains. "There's no more copy-and-paste, or jagged logos." In today's attacks, visual components of brand impersonation almost exactly mimic true content.

This presents a clear security hurdle, he continues, because people and technology can no longer look for artifacts that previously distinguished fake content from the real thing. "Those visual cues are not there anymore," says Grana of a key challenge the research team faced.

Most people are familiar with the concept of image recognition. What makes detecting brand impersonation different is twofold: For one, a victim may receive different types of content that aim to imitate the same brand. An impersonation attack spoofing Microsoft, for example, might send one malicious email that mimics Excel, and another designed to look like Word.

"Those are two very different pieces of content, even though they both represent Microsoft," Grana says.

While too many types of content can present a detection challenge, too few can do the same. Many brands, such as regional banks and other small organizations, aren't often seen in brand impersonation, so there might only be a handful of training examples for a system to learn from.

"The standard deep learning that requires tons and tons of examples per class – class is the brand in this case – really wouldn't work in our situation," he notes.

To address the issue of detecting brand impersonation attacks, Grana teamed up with software engineer Yuchao Dai, software architect Nitin Kumar Goel, and senior applied researcher Jugal Parikh. Together, they developed and trained a Siamese Neural Network on labeled images to detect these types of attacks. Unlike standard deep learning, which is trained on many examples, Siamese Neural Networks are designed to generate better predictions using a smaller number of samples.

[The researchers will discuss their approach, further applications, and planned improvements in their upcoming Black Hat briefing, "Siamese Neural Networks for Detecting Brand Impersonation" on Wednesday, Aug. 4]

The team's dataset consists of more than 50,000 screenshots of malicious login pages spanning more than 1,000 brand impersonations. Each image is a collection of numbers, Grana says, and the team translated those numbers into what he describes a "point" on an N-dimensional coordinate plane. Instead of an image, which has three dimensions of all its different pixels, it becomes numbers. The team sought a way to make the numbers meaningful and in doing so, distinguish fake from real brand images.

"Our algorithm that we used, we rewarded it for … translating content of the same brand to similar numbers, and contents of different brands to different numbers, so that way, when we look at these new numbers that are now meaningful because we trained our network to do so, any numbers that were close together were likely from the same brand," he explains.

Their Siamese Neural Network learns to embed images of the same brand relatively close together in a low-dimensional space, while images of different brands are embedded further apart. They then do a "nearest neighbor classification" in the embedded space.

Training Models, Learning Lessons
Grana says the team faced quite a few challenges and learned some lessons along the way.

"Dealing with skewed data is a large issue," he notes. "When you have a dataset that only has a couple observations per brand or per class, it really does require special techniques. We did some testing with the normal neural network, and it just wasn't sufficient for our purposes."

Determining the specific techniques that will work requires a lot of trial and error, Grana says of the research process. Which method will best suit the data you have? "There's the science behind machine learning, but there is also the art of it, to say, 'which optimization algorithm should we try; which network architecture should we try,'" he explains.

The researchers' work is still ongoing, he adds. Their next goal is to examine how this approach might work with a smart and adaptive adversary, as a means of improving the technology and response to attackers' evolving techniques. The screenshots they used in this research won't be the same ones used in future attacks, and security tech needs to keep pace.