Advertising Algorithms Could Be the Solution to Gaining Communication Network Intelligence
The same technology that allows online advertisers to target consumers as we browse the web could eventually play a key role in communication network security – even at the national-defense level.
An award-winning group of researchers published a study in IEEE Xplore saying that advertising algorithms—frequently used by search engines like Google and Yahoo—could help develop a machine-learning approach for navigating complex communication networks. In the past, it was relatively easy to both protect communication networks from attempted hacking and intercept (or jam) signals from other networks.
These days, though, networks are anything but simple thanks to various factors such as mobile devices, sensing technologies, social networks, and others. As a result, understanding how these networks communicate is a challenge. This also makes developing any kind of machine-learning approach to gain artificial intelligence a challenge as well, since that depends on being able to track and identify patterns.
The below figures illustrate the differences between older communication networks (left), and today’s much more complex networks (right). In the past, it was common for there to be a main entity which, when attacked, would break the whole system. But with the advent of more sophisticated channels and devices, communication topologies are increasingly decentralized and without clear patterns.
This makes it difficult for military operations to perform network interdiction, or jamming, to gain information about the network via monitoring and interference. Ultimately, network interdiction can stop malicious attacks and provide key leads on enemies, but it’s also a tactic that can be used defensively to build robust networks and thereby protect it from potential eavesdroppers. This concept can also be applied in commercial settings.
SaiDhiraj Amuru, chief engineer at Samsung R&D Institute, along with a team of researchers examined advertising algorithms used to optimize online banners, which are based on a classic concept called the multi-armed bandit, or exploration-exploitation, dilemma. In simple terms, this dilemma is what tells the system that certain ads work better than others at getting clicks—or website visits—from specific users. It then encourages the system to use the same or similar ads in the future to target those users for a certain amount of time. Thus, a pattern is identified.
The same concept can be applied for blind network interdiction.
Consider, for example, a network that has 10 communication sources. An outside government entity, “Joe,” attacks one source to see how the network reacts. If the other sources continue acknowledging the source, chances are the source under attack was not the primary target or leader. If Joe continues attacking the same source every time with no substantial change in network acknowledgement, it’s time to move onto another source. Once an attack leads to reduced or no communication acknowledgment from the network, the lead on the enemy target becomes stronger. All along, Joe is gathering feedback and learning the topology.
Amuru and his team refined the advertising algorithms to give them boundaries on when to keep trying, and when to attack a different node over a certain period of time. They found that when tested against more popular learning algorithms such as what’s called upper confidence bounds (UCB) and contextual bandits, their blind learning algorithm successfully attacked more messages.
“If adopted, this approach could lead to quicker ways to navigate complicated communication networks with very little information to start with,” said Amuru. “However, we’re still addressing some challenges to try to expedite the time it takes to collect feedback. We’re also focused on finding ways to rank messages based on their critical importance to further improve model efficiency.”
The team is currently looking into extending the algorithms to scenarios with multiple attackers, as well attacking just one person vs. an entire network.