\
Case study
\
Case Study
How Reality Defender Exposed a Political Disinformation Campaign in Taiwan
As deepfakes and AI-generated content become more sophisticated and accessible, the bad actors using these tools to sow disinformation acquire more influence. Employing a range of methods, from off-the-shelf technology for creating "cheapfakes" (low-to-mid-quality deepfakes) to resource-intensive yet high-quality audio manipulations, politically motivated culprits are increasingly disrupting the news cycle.
Faced with convincing deepfaked voice clones, news and social platforms further amplify the dissemination of media depicting politicians and prominent figures uttering statements they never actually made. This phenomenon is exemplified by the case of Ko Wen-Je, former mayor of Taipei and the official nominee of Taiwan People's Party (TPP), who became the target of a recent disinformation campaign. The incident was centered around a deepfake audio clip propagated by media outlets in Asia.
The Challenges
Ko Wen-Je, a major candidate in the high-stakes 2024 presidential election in Taiwan, claimed to be the victim of an audio deepfake in which his cloned voice can be heard disparaging one of his political opponents. According to several state media outlets, the 57-second audio file was emailed to journalists in 2023, and the clip was spread by those same outlets without verification of its authenticity. In the deepfake clip, the candidate's voice is made to assert that his opponent, presidential nominee for the Democratic Progressive Party (DPP) and current vice president, Lai Ching-Te, sought an inappropriate relationship with the United States government and paid people to attend his political event, among other criticisms and insults.
In what was one of the highest-profile examples of election meddling via voice cloning to date, Ko Wen-Je's reputation was seriously damaged, and the audio succeeded in its creator's objective of sowing conflict and confusion among the country's electorate.
A voice clone deepfake can be created without much effort, and because the candidate is a public figure with regular media and online presence, recordings of his likeness and voice can be easily obtained. With merely a few seconds of the candidate's speeches, malicious actors could clone his voice with publicly available tools and a few seconds of their time.
The video was then widely disseminated with the help of state media figures in China, a country that had an interest in the outcome of the election.
Following the release of the audio clip, Ko Wen-Je and his political party immediately sprung into action. At a news conference, the candidate condemned this attempt at defamation and sowing conflict and confusion among Taiwan's voters. The political party submitted the email containing the deepfake audio file to the country's Ministry of Justice. Fortunately, its Investigation Bureau relied on Reality Defender to aid in their inquiry.
The Solutions
After consulting with Reality Defender, Taiwan's Investigation Bureau took advantage of their access to our best-in-class deepfake detection platform to submit the audio clip in question for analysis. After a simple upload, Reality Defender's proprietary detection tools scanned the clip and immediately provided a result: the audio was overwhelmingly likely to be a deepfake.
To address the complexity of AI-generated content, Reality Defender's platform refrains from offering a simple "yes" or "no" answer regarding the analyzed media. Instead, we furnish a probability score (ranging from 1 to 99) in a comprehensive report on the submitted content. This approach allows the client to review the results and draw conclusions independently.
Our models train on thousands of existing deepfake models while also projecting for future deepfake creation methods and model types to ascertain the authenticity of media. This enables Reality Defender to not only furnish up-to-date detection on the newest and most current deepfake methods, but also to stay one step ahead of emerging models that may arise. With our tools, Taiwan's investigative authorities remain prepared for attempts at election meddling via generative AI technology.
The Advantages
In this instance, the audio submitted by the Ministry of Justice was given a high score by our detection models, indicating that the clip was highly likely to be a deepfake.
This enabled Taiwan's authorities to initiate a more thorough investigation into the incident, aiming to identify the malicious actor(s) responsible and, if found, prosecute them to the fullest extent of the law.
The voice-cloned presidential candidate and Taiwan People's Party were able to apprise the electorate about the incident, restore their public image, and advocate for a fair and free election process. Simultaneously, they were able to demonstrate that the defamatory statements about Lai Ching-Te were a maliciously synthesized fiction.
Given the sensitive nature of Reality Defender's role in uncovering political deepfakes and safeguarding the integrity of elections, the company's CEO, Ben Colman, was unable to confirm or provide specific comments on our collaboration with the Investigation Bureau at the time, despite numerous inquiries from international media outlets.
Reality Defender's clients can rely on absolute security and anonymity regarding their use of our technology.
The company does not disclose specific clients, and the media they scan—along with their reasons for scanning—remains entirely inaccessible to our team due to the company's strict security and privacy protocols. We are able to acknowledge our partnership at this time because Taiwan's Ministry of Justice revealed publicly their usage of Reality Defender's detection tools in the case.
Conclusions
In the months since this incident, there were additional attempts to sow confusion and chaos in Taiwan's presidential elections, and the country's authorities continued to depend on Reality Defender's deepfake detection technology to safeguard the integrity of its democratic process.
Due to widespread dissemination and promotion of the initial audio clip among state-owned media outlets before Taiwan's authorities initiated their investigation, the revelation that the media files were deepfakes garnered less attention than the news of the initial clip.
This incident represents an ongoing trend of manipulating news media with strategically placed deepfakes, as the content of these deepfakes tends to spread more extensively than the disclosure of their inauthenticity. This underscores the urgent need for media institutions, social media platforms, and government agencies to scan for deepfakes at the outset, preventing potential harm.
Such an implementation would empower media outlets to enhance their research and fact-checking capabilities, while enabling platforms and government agencies to identify bad actors before they can inflict irreparable harm and sow discord.
In the end, Taiwan's 2024 presidential election did not appear to be materially impacted by deepfakes, and Lai Ching-Te is the new President-Elect. The country's criminal justice institutions continue to rely on Reality Defender to refute (or verify) media used to present synthetically created statements post-election, especially if post-electoral processes are equally bombarded by falsified media. As we persist in addressing newer deepfake models and fortifying our detection algorithms against upcoming challenges, we stand prepared to assist public servants in countering misinformation and thwarting the bad actors targeting them.
Get started with Reality Defender's deepfake detection platform today.
\
Case Studies