Deficiencies of Fake News Automatic Detection System

2019-10-18

The 21st century is filled with a large amount of false information on the internet, which is an inevitable problem in the era of big data. Entertainment information is not to be mentioned. According to publicly available research data from the European Research Council, during the 2016 US presidential election, one in four Americans had at least browsed through a fake news article. Subsequently, with the emergence of automatic text generators, the problem of fake news has become even more serious. As a result, researchers have developed an automatic detector that can identify machine-generated text. However, a team from the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT has found that this method is not perfect. In order to prove their point, the researchers from this team developed an attack system that can deceive the most advanced fake news detector. This system mainly relies on imitating human writing to deceive the detector. Besides being less flexible, the detector itself has a major flaw: most of the time, it will consider machine-generated text as fake, so it may mistakenly classify correctly and legally generated text as fake news. To address this issue, the researchers from MIT proposed a strategy to help the detector overcome its shortcomings. The team created a dataset that eliminates inherent biases and changes the machine's thinking pattern to correct some of the detector's errors. They then developed a new algorithm that reduces the importance of phrases that are particularly common in their respective categories, while increasing the importance of phrases that are rarely seen in that category. In simple terms, by changing the weight of phrases, they allow the machine to break free from its inherent thinking pattern and fundamentally restrain the automatic text generator that generates articles based on common sense. Using this approach, the current text generator can be effectively restrained, but it is only a temporary solution. As long as someone with malicious intent changes the algorithm sequence of the text generator, they can in turn restrain the detector, creating an almost unsolvable loop. In response to this, the team proposed a strategy to combine the verification of truth and falsehood with existing defenses, making the model more sensitive to attacks. At the same time, they aim to further improve the existing model by developing new algorithms and constructing a dataset that covers more types of false information. With the determination to solve the problem from its root, the research team from MIT will submit a research report based on bias removal and truth preservation at the "Empirical Methods in Natural Language Processing" conference in Hong Kong in October, publicly sharing their research findings.

RECENT AI NEWS

RECENT AI TOOLS