Award on the international stage
The research work of Pietro Melzi and Christian Rathgeb on the topic of anti-bias in models with artificial intelligence (AI) has won the "IEEE AFMG International Workshop Best Paper Award" at the International Conference on Computer Vision 2023 (ICCV 2023). The award is a significant recognition of secunet's research work in the field of AI and also highlights the importance and relevance of the work in an international context.
The ICCV, one of the leading conferences in the field of computer vision, provides a platform for experts from around the world to present their latest research findings and technological breakthroughs. It is considered the most prestigious event in this sector.
The publication addresses the minimization of demographic biases in AI models in the context of image analysis. At a time when the fairness and impartiality of AI systems are increasingly debated, our research offers innovative methods to address these challenges.
The research addresses the problem of statistical distortions, also known as bias, in AI models. These biases can arise if the data used to train the AI does not represent all features equally. This can lead to certain groups of people being favored or disadvantaged and also jeopardize the functionality of the models.
The secunet Anti Bias approach comprises a three-stage solution:
This innovative method not only generates measurable fairness and non-discrimination of the AI models, but also increases their security and robustness. In this way, secunet is addressing the challenges of modern image processing and ensuring greater trust in AI technologies.
We iIdentify and correct biases related to characteristics such as age, gender and ethnicity are identified and corrected in the early stages.
By generating photorealistic artificial identities that vary in different characteristics, the recognition performance of the AI model is tested. This enables a comprehensive check for possible bias.
If bias is detected, new identities are generated for the training data and the model is re-trained until no more bias can be detected.
"The success at ICCV 2023 is a significant milestone for our AI research. We are proud to continue on this path and develop solutions that are not only technically advanced, but also ethically responsible."
