Two Swiss startups take digital trust to the next level

PXL Vision and Idiap develop solution for detecting deepfakes

PXL Vision, the leading provider of digital identity verification solutions, and the Swiss research institute Idiap are joining forces to combat fraudulent AI-generated facial images and travel documents. The project, funded by the Swiss innovation agency Innosuisse, aims to bring the world's first robust solution for deep facial recognition to market. "Identity forgery using artificially generated data is a growing problem in the fight against attempted fraud on the Internet. Together with Idiap, we are therefore working at full speed on the next generation of deepfake detection to adapt our digital identity verification solutions to the new technological challenges and make them even more secure," says Michael Born, CEO of PXL Vision. "By generating synthetic data for document and image recognition, we also want to ensure that our facial verification always has the same high quality, regardless of the age, gender or skin colour of the people depicted."

General deepfake detection is still a technical challenge even for the most modern technology. "With the rapid development of artificial intelligence methods, we are striving to stay ahead of the curve. Together with Idiap, we want to develop a solution that can be used with novel methods to recognise both AI-generated faces and identity documents. The same technologies will be used to cover the entire global population with a new synthetic and balanced dataset of faces (in terms of gender, age, skin colour) to ensure the accuracy of facial recognition by PXL Vision products in the future." adds the CEO of the Swiss Identification company.

Tune Insight integrates LLMs and collaborates with EPFL

LLMs (Large Language Models) have recently become increasingly interesting tools since the arrival of OpenAI’s chatGPT – a chatbot based on GPT – that reached 100 million registered accounts in less than 3 months. At the forefront of cutting-edge technology with an unwavering commitment to data security and privacy, Tune Insight announced initial support of LLMs into its product to enhance user-experience. In healthcare, Tune Insight’s federated health data space solution now integrates LLMs to facilitate the generation of complex database queries for cohort exploration, and to provide users with additional knowledge around collective results computed by the platform. When those features are enabled, the used LLMs can be hosted on the Tune Insight instance at the customer, or in another trusted or sovereign cloud environment. In addition, Retrieval-Augmented Generation (RAG) is used to enhance the quality of the outputs using confidential internal documents, while guaranteeing that no private information ever leaves the controlled customer environment.

In addition, through a project recently awarded by Switzerland’s innovation agency Innosuisse, Tune Insight and the EPFL Machine Learning and Optimization lab led by Prof. Martin Jaggi join forces to forge the new generation of privacy-preserving and federated LLMs. This collaboration includes the exploration of secure federated fine-tuning of medical models such as Meditron, EPFL's new LLM for medical knowledge.

The envisioned privacy-preserving and federated capabilities will enable companies to use and benefit from LLMs with data protection and privacy guarantees, relying on best-in-class models while preventing the disclosure of usage statistics with other companies, and subsequently being able to refine existing models with their own data, even the most sensitive data, without disclosing it.

(ES)