Researchers Adapt AI With Aim to Identify Anonymous Authors

With disinformation on social media a significant problem, the ability to identify authors of malicious articles and the originators of disinformation campaigns could help reduce the threat from such information attacks.  

At the Black Hat Asia 2020 conference this week, three researchers from Baidu Security, the cybersecurity division of the Chinese technology giant Baidu, presented their approach to identifying authors based on machine learning techniques, such as neural networks. The researchers used 130,000 articles by more than 3,600 authors scraped from eight websites to train a neural network that could identify an author from a group of five possible writers 93% of the time and identify an author from a group of 2,000 possible writers 27% of the time.

While the results are not impressive, they do show that identifying the person behind a piece of writing is possible, said Li Yiping, a researcher at Baidu Security, during his presentation on his team's work.

"Most fake news is posted anonymously and lacks valid information to identify the author," he said. "Tracking anonymous articles is a challenging problem, but fortunately it is not impossible. Different people have different writing styles, so we are able to identify some writers by their distinct habits." 

Fake news and other forms of disinformation have become an online plague over the past decade. Driven by commercial success, cybercriminals have used fake news to attract page views against which advertising is sold. More insidious, however, are political disinformation campaigns by foreign nations and domestic groups with agendas that can impact public opinion using untrue information. 

In late September, the FBI and the US Department of Homeland Security issued a warning that both foreign actors and cybercriminals will likely use disinformation in various campaigns this election season.

"Foreign actors and cybercriminals could create new websites, change existing websites, and create or share corresponding social media content to spread false information in an attempt to discredit the electoral process and undermine confidence in U.S. democratic institutions," the agencies stated.

A variety of research efforts are underway, aiming to unmask disinformation campaigns. In May, for example, a group of of researchers at NortonLifelock launched BotSight, a plug-in that rates social media accounts on a bot-versus-human scale. The tool uses the known connections between social media accounts to calculate a probability that a specific account is managed by an automated bot or an actual human.

At the Black Hat USA conference, a research manager at the Stanford Internet Observatory argued that Russia tends to focus more on disinformation campaigns involving fake memes and articles, while Chinese efforts focus more on creating legitimate-seeming news sources that espouse a government-approved focus.

Baidu Security's research effort focused on either matching an article to a known author in a list of sources, called the author attribution problem, or determining the likelihood that an article was written by a specific author, known as the author verification problem. The researcher trained a neural network using a series of triplets of article data: an anchor article written by an author, an article that positively matches the author, and an article that was not written by the author. 

By using a dynamic method of selecting such only a small share of possible triplets, the research team created a training data set to create a neural network that identifies the author of an article. In an experiment using seven datasets of increasingly complexity, the researchers found their method worked well, with 93% accuracy, in attributing any of 600 articles written by five different authors, but was only 27% successful in attributing more than 70,000 documents written by any of 2,000 different authors. 

Researcher Li noted that, even at such a low accuracy with a high number of documents, the Baidu team's approach had better accuracy than other methods.

"Our method outperformed other baselines, especially when the data sets get large," he said. "In the future, we will continue to test our model and optimize our deep learning network and triplet selection strategy."