INTA 2019: Anti-counterfeiting is like “playing chess while the rules are changing”
Google and Amazon trademark counsel weighed in on their approaches to online infringers and social media, with Google’s lawyer explaining how effective AI is in the fight against counterfeits
Amazon’s senior trademark counsel Stephen Coates, who was speaking on yesterday’s panel titled “Combating Counterfeiting on the Internet Highway and in Digital Media” during INTA’s Annual Meeting in Boston, said that counterfeiting through social media has substantially increased in the past three years.
“At the time there were very few counterfeit complaints,” he said, reflecting on a talk he gave in 2016. “Oftentimes they would be in the single or double digits – mostly less than 20.”
He stressed that back then he did not see counterfeiting as a problem on the platform of Twitter, his employer at the time. Now on this year’s panel, Coates said: “Anecdotally, I see an increase in the use of advertising on social media to facilitate the sale of counterfeit products.”
Coates’s role at Amazon involves protecting products in the company’s digital device arm, including Kindle, Echo and the recently acquired Ring doorbell units. He demonstrated various ways the company’s Ring device has been infringed by counterfeiters, including through Instagram accounts, Shopify stores and Facebook groups. In the instance of one Shopify account, the infringer was not only selling a knockoff product, but was charging consumers a 40% markup.
“In these cases you’re faced with the choice of doing a test buy, and for online counterfeiters I find that test buys are just too slow – especially when the counterfeiter or the sale looks obviously illegitimate,” Coates said. “It’s often better to go down the road of a takedown to get that stuff removed.”
Coates then described how apps themselves have been vehicles for infringement or fraudulent activity.
“We’re seeing an increase in apps that are used to interact with devices, including our own,” Coates said. “That obviously could be a concern for us. The potential risk is the app developer could be collecting user login information, or accessing their accounts, or worse yet, they could potentially be used to co-opt users’ devices.”
Some infringing apps have been found to contain malware as part of the download, Coates said.
Machine learning curve
Google’s legal director for trademarks, Annabelle DanielVarda, spoke about artificial intelligence as a remedy for online infringement.
In her position, a key area where counterfeiters are targeted is Google Ads, the company’s main platform for online advertising and sponsored listings. Her group’s anti-counterfeiting tactics rely heavily on machine learning to identify and remove advertisers that are selling infringing products.
Engineers at the company have developed risk models that look for signals which identify an advertiser as a “bad actor,” DanielVarda said. “We find that these bad actors tend to use the same techniques, irrespective of abuse area that we specialise in – and counterfeiting is just one area.”
DanielVarda described some of the signals that these risk models look for, including: 1) advertisers using a proxy IP address to avoid being traced online, 2) using zeroes instead of Os in names or titles to evade search detection, 3) being active at unusual times for the advertiser’s supposed location, and 4) there being logins from geographic areas with a pattern of infringement activity.
One of the advantages of machine learning is its ability to cluster data and detect patterns, DanielVarda said. “These models are not just looking at all these different things independently,” she said. “Models are able to associate different signals that individually may seem innocuous, but when seen all together, can help identify a bad actor.”
To anyone asking how effective machine learning models are as anti-counterfeiting measures, DanielVarda said: “In short, they’re very effective.”
She said that more than 99% of the accounts that Google terminates on counterfeiting grounds are done proactively through the risk models. In 2011, when the group started relying on machine learning, the risk models terminated more than 150,000 accounts. Last year, they dropped to 10,000, showing a strong deterrent effect.
In addition, the risk models’ ability to cluster information has led to subsequent flags or connected accounts after one infringer is identified, helping to prevent the dreaded “whack-a-mole” problem.
“The original engineer who led the development of these risk models always says that combating bad actors is like playing chess while the rules are changing in the middle of the game,” DanielVarda said.
Coates described how tightening privacy laws have presented difficulties in combating online infringers. “The implementation of GDPR means that information is slowly being erased around the world – information we would use to identify who potentially is operating a website,” he said.
However, ISPs have been more willing to provide information on website registrants that is privacy-protected, Coates said, adding that providers including Cloudflare and GoDaddy have been active in turning over information for anti-counterfeiting purposes.