Presentations at the Networks and Distributed System Security (NDSS) Symposium 2024

The System Security Lab was proud to contribute four papers focusing on the security of Federated Learning and the detection of AI-generated texts. On February 28th, Phillip Rieger introduced “CrowdGuard,” a novel methodology for detecting backdoor attacks within Federated Learning frameworks. Unlike traditional server-side defenses, which are limited to comparing model updates, CrowdGuard employs a novel architecture that allows to utilize clients’ data for validation purposes. Through the use of secure enclaves on the client side, CrowdGuard examines changes in model behavior to identify malicious updates effectively.

On the same day Alessandro Pegoraro presented “FreqFed,” an innovative approach to counteract poisoning attacks in Federated Learning. By analyzing model updates in the frequency domain, FreqFed leverages the insight that new behaviors predominantly influence the low frequencies of transformed models. This technique enables the precise identificationand mitigation of manipulated models.

On February 30th, Kavita Kumari showed our latest efforts in distinguishing AI-generated texts from those written by humans. As tools like ChatGPT become increasingly prevalent, sophisticated methods are essential for accurate detection. Our approach, “DEMASQ,” leverages the principles of the Doppler Effect and energy-based models to identify artificially generated texts.

The symposium served as a fruitful platform for engaging discussions on our current research efforts. It facilitated valuable exchanges with the broader security community, fostering an environment conducive to collaboration and innovation.