In the recently held 2022 World Artificial Intelligence Conference, the WAIC 2022 Data Element Circulation Technology Frontier Exploration Forum was one of the major theme forums of the conference. With the theme of "Open Symbiosis, Integration of Data and Reality", the forum focused on the important economic and strategic value of data as a key production factor driving economic and social innovation and development, as well as the corresponding security threats and privacy challenges.

AI Data Security and Hidden Dangers

Zhou Aoying, vice president of East China Normal University, said at the forum, "We are at the best time for database development, with rich practical scenarios, we can uphold abstract concepts and establish new technical and theoretical systems." However, in the era of artificial intelligence, how do we protect personal privacy?

For example, the current AI synthesis technology is widely used in speech, image, video, and other fields. At present, it has been possible to easily make fake face-changing videos through AI technology. If it can be better developed and utilized, it will greatly reduce the cost of film and television editing and the development cost of film and television dramas. However, if a personal portrait is used in a video without the personal knowledge, it will inevitably involve infringement of personal portrait rights, reputation rights and other legitimate rights and interests. In addition, the current speech synthesis and speech conversion technology has also been able to mix the fake with the real and can extract the speaker's voice representation through the voiceprint model to synthesize the voice and use it for voiceprint unlocking.

These deception techniques are called Spoofing Attacks. Even if a little random noise disturbance is added to the real image, the image will be recognized as other content. Such as the panda image in the figure below, after adding some random noise disturbance, it will be recognized as a gibbon. Then for the category of gibbons, the perturbed panda picture is an example of an adversarial attack, which will confuse the ability of the AI recognition model. So, how should we deal with the above-mentioned AI model system adversarial attacks and AI data security issues?

Anti-Adversarial Attacks on AI Models

At present, there is a special academic research topic - Anti-Spoofing. This topic is specially designed to detect AI model recognition, AI synthesis or artificial synthesis data to determine its authenticity. First, add a true and false judgment module to the picture, voice or video you want to recognize, and then give the authentication result, whether it passes or not. The figure below shows that an AASIST module is added before the speaker is confirmed, that is, the speech authenticity detection module. This module is integrated with the voiceprint recognition module, and then it is determined whether the two voices belong to the same speaker.

AI Data Security

Regarding the security of AI data, China has recently introduced a series of laws to ensure the security of private data. For example, the Personal Information Protection Law and the Data Security Law both refer to the protection of personal and corporate data. This is a new regulation that will have a far-reaching impact on the future of the entire artificial intelligence industry, and it is also a new requirement to comply with the development trend of the industry. But these are far from enough.

As an individual, you also need to face the phenomenon of big data killing. If the community or school collects personal face images, voice, video, and other data without the permission of the image collector, it is actually an infringement of personal privacy data. Formal data collection needs to rely on professional data companies, which not only comply with regulations in protecting user privacy, but also provide more accurate and high-quality data.

As a professional data company, Magic Data always putting data security at the first priority, designing and applying a strict data protection mechanism so as to provide trusted AI training data for the industry. The internal processes are in accordance with industry security standards, and are GDPR compliant, ISO 27001 and ISO/IEC 27701:2019 certified.

Data compliance are strategically incorporated into Magic Data’s development and running through management and technician level. Magic Data organizes regular trainings on data security and compliance and regularly strengthen data governance, management and compliance through third-party legal and technical advisory consulting, assessment and audit services.

We are proud to be compliant with GDPR and PIPL and accredited with a range of standards and certifications including ISO/IEC 27701:2019 and ISO27001.

Contact us by emailing business@magicdatatech.com for more information.