Monitoring wireless networks is much more challenging than monitoring those that are wired, says David Wetherall, director of Intel Research Seattle. In wired networks, he explains, one can attach a box to the network hardware that will reliably count the number of packets in and out. But in a wireless network, it’s impossible to collect all the packets, and these lost bits invariably obscure the picture of actual activity. To solve this problem, the UCSD researchers developed a novel set of algorithms that infer wireless activity that isn’t directly measured. For instance, if a monitoring radio sees that a laptop received a packet, but didn’t see that a packet was sent to the laptop, the algorithm can infer that a packet was sent. The researchers’ model, explains Savage, infers behavior about activity at many different levels of the network. The model takes into account the structure of the underlying wired network, the method used to encode data into wireless signals, and the manner in which access points and wireless cards in laptops send and receive information.
The researchers have “done a nice job extending and synthesizing known inference techniques into a useful system,” says Wetherall. He suspects that it wouldn’t take much to use this approach to make a commercial system.
While the approach couldn’t solve the problem of spotty wireless coverage in an apartment building where access points aren’t connected to a single wired network, it could be modified to monitor citywide Wi-Fi. Savage says that his team is looking at better ways to monitor large-area Wi-Fi, which is more difficult to do than monitoring Wi-Fi in a building. “If you’re trying to cover a large area, then there may be some places that you don’t see at all, or you just might have one data point for.” In this case, he says, the inference algorithms would have to be tweaked to make more guesses based on much less information.