Deepfake is a technique that uses deep learning and artificial intelligence technology to create highly realistic illusions, capable of simulating human voice, images, and video among other forms of content. Recent surveys have shown that with the rapid development of AI technology, illicit activities using AI-related technology are significantly on the rise in the dark industry.

In the past few years, we have witnessed events such as a 2019 incident where AI voice imitation software was used to fake the voice of a top executive and swindle 220,000 euros from a UK energy company's CEO. In the era of large models, similar deep fake actions are expected to become more frequent.

Xinye Technology recognized this challenge and thus recently launched the ninth "Xinye Technology Cup" Global Artificial Intelligence Algorithm Competition, aimed at seeking and promoting breakthroughs in voice deepfake detection technology.

The Vice President and head of Big Data and AI, Chen Lei expressed that they hope AI voice anti-fraud technology can work in coordination with large model application technology, forming an effective countermeasure mechanism to better protect citizens' information security and rights.

Especially in high-value fields such as fintech, deepfake crimes can cause even more serious damage. The report points out that during 2022 to 2023, related identity fraud cases increased significantly, with the Philippines as an example, where the number of scam attempts increased by 4500% year-on-year.

Chen Lei noted that despite the rapid development of voice synthesis technology, it has been relatively lagging due to the one-dimensional continuous signal in voice containing complex information such as accents, intonations, and dialects.

To more effectively face the challenge of fake voices, Xinye Technology has used its capabilities in voice synthesis, voice recognition, and voiceprint recognition technology to provide risk assistance analysis functions for financial services. Through hosting global AI algorithm contests, they gather AI experts and enthusiasts worldwide to jointly crack the challenge of voice deepfake detection.

Xinye Technology's algorithm scientist Lu Qiang explained that this competition will use the widely applied voice processing technology and the latest text-to-speech (TTS) models to generate data, construct datasets for the competition, and deliberately add fake voices generated by the latest large models in the semifinal stage to make the competition more challenging and unique.

"In today's competition, we expect to see contestants come up with more efficient methods to distinguish new kinds of forged voices, especially those generated by advanced large models." We have now entered the era of large models, where various technical fields are highly reliant on the support of large models.

Traditional techniques seem powerless when facing the complex task of identifying synthesized fake speech generated by large models. These traditional models tend to overlook many details when handling large amounts of data, whereas large models can easily tackle this challenge. It is for this reason that the technology for the detection of forged speech urgently needs to be upgraded. Only by utilizing powerful tools like large models can we more accurately capture the subtle features of fake speech.

Typically, physiological changes occur when people lie or experience emotional fluctuations, and lie detectors work by capturing these changes to detect deceit. Similarly, the working principles of AI large models in identifying fake speech are akin to those of lie detectors, modeling the precise capability to capture physiological responses, thus identifying those forged minute fragments among a vast amount of genuine data, connecting these fragments, and eventually judging the authenticity of the speech through quantified data.

In this process, the fake speech detection model acts like a shield, aiming to strengthen data security and protect user privacy.

To more effectively detect fake speech, Trust Also Technology is dedicated to constructing datasets of authentic and fake speech, continually enriching and perfecting the dark industry database, and enhancing the capability to combat black market industries and fraudulent activities by iterating on powerful models. The company also plans to actively adopt an open-source philosophy, build open-source datasets with desensitized data, and promote collaboration between industry, academia, and research. After the conclusion of the "Trust Also Technology Cup" competition, the accumulated competition data will also be open-sourced.

It is worth noting that we live in an era where everything is connected, and voice anti-counterfeiting should not be limited to a single modality. With technological advancement, we can use not only the voice itself for identification but also combine video, text, and other multimodal data in the future to improve recognition accuracy. This is particularly true in the field of financial technology, where voice identification of fraudulent behavior is inseparable from actual application scenarios.

Currently, Trust Also Technology has collaborated with public security systems in regions such as Shanghai and Fujian, sharing dark industry databases, to assist in identifying black market behaviors. Facing the increasing number of fraud activities induced by false information, the AI anti-counterfeiting capabilities built domestically can be transferred to international business in the future, offering support for financial technology companies as they enter international markets. The company has already launched abroad algorithms for identifying fake faces, fake documents, and voice recognition. By integrating voiceprint services in apps, combined with face recognition and voiceprint identification technologies, Trust Also Technology can effectively identify illegal impersonators, assisting risk management teams.