WUT representatives on the podium at the CISPA hackathon

AI models put to the test

Over the course of 24 hours, participants worked simultaneously on three advanced challenges in the field of Trustworthy AI. The tasks focused on artificial intelligence and cybersecurity, specifically the privacy and trustworthiness of machine learning models. The final ranking was based on the combined score achieved across all three competition tasks, as well as the teams’ presentation of their results and solution concepts before the jury. 

The first challenge involved identifying watermarks in text. For each token in a document, participants were required to estimate the probability that it had been generated by a language model with an active watermarking mechanism. The documents contained a mixture of human-written text, unwatermarked AI-generated text, and passages watermarked using several different techniques. The second challenge, Member vs Generated Inference, required participants to modify images in such a way that a hidden detector would incorrectly classify them as belonging to a different category—training data, non-training data, or AI-generated data. The final challenge focused on reconstructing data from gradients in federated learning. Participants were required to recover 1,536 private training images using only the gradients exchanged between the client and the server. The task demonstrated that federated learning, promoted as a method for protecting privacy, does not guarantee such protection on its own. 

As Bartłomiej Dmitruk from the Advanced Persistent Thinkers team explains, the greatest challenge was the time pressure combined with the high level of difficulty of the tasks. The solutions had to work on previously unseen data, and the problems were highly demanding—for example, detecting watermarks while maintaining an extremely low acceptable false-positive rate. Our student also emphasizes that participating in the competition provided the team with valuable experience.

“What benefited us the most was meeting top-level specialists: researchers from CISPA and the best teams from across Europe. We also had the opportunity to work on cutting-edge topics at the intersection of security and AI—topics that are usually encountered through research publications, but here we had to implement them ourselves under time pressure—as well as to develop skills in rapid prototyping and teamwork. What we liked most was that the tasks were based on real research problems,” says Bartłomiej Dmitruk, a student at the Faculty of Electronics and Information Technology at WUT. “The organization was excellent, the scientific level was very high, and the final event, including the gala at Ingobertusfest, was truly exceptional. We plan to participate in more hackathons of this kind in the future.”

Advanced Persistent Thinkers team awarded at the CISPA Hackathon

Aleksander Ogonowski from the Syntax Terror team also highlights the time pressure, as well as the highly competitive nature of the event. The leading teams were very evenly matched, and the final ranking remained uncertain almost until the last submissions were made.

“The tasks differed from what we deal with on a daily basis, and 24 hours is not much time to understand the problem, test several approaches, and refine the solution. There was also the challenge of organizing the work within a four-person team: we had to quickly divide responsibilities, work in parallel on three independent problems, and continuously decide where to invest the remaining time. Another challenge was bringing our solutions together into a short presentation,” says Aleksander Ogonowski, a doctoral student at the WUT Doctoral School. “The tasks addressed topics that are currently the subject of active research, including attacks on model privacy, adversarial attacks, and the detection of AI-generated content. Working on these problems at such a pace requires quickly engaging with new research literature and empirically validating hypotheses rather than developing a single idea in depth. This is a skill that directly translates into research work. The organization was at a very high level - the efficient access to GPU resources provided to all teams deserves particular recognition. The tasks were well designed and strongly aligned with current technology trends, which made working on them an interesting experience from our perspective.”

Syntax Terror team awarded at the CISPA Hackathon