The security of digital systems is under constant threat of attacks. One way to improve cybersecurity is to predict how hackers could manipulate new technologies to break into existing systems. However, little is known about how cybercriminals might take advantage of the emerging field of machine learning. Funded by the European Research Council, the MALFOY project aims to determine how machine learning algorithms can be used to discover security weaknesses and perform computer attacks automatically. By taking the position of the attacker to explore offensive security techniques, the project will be able to construct effective defense mechanisms.

ERC project conducted in the TU Berlin MLSec group

International bypass frauds, also known as SIMBox frauds, rank among the most prevalent scams in cellular networks, placing in the top three phone system frauds. This fraud results in an annual global loss of $3.11 billion and threatens network users’ privacy and national security. Despite its widespread impact, SIMBox fraud remains a persistent challenge, continuously evading detection due to the continual refinement of fraudulent behavior. This project leverages ML through an analysis and formalization approach to understand the evolving fraud behavior in diverse cellular network traces, including Charging Data Records (CDRs) and signaling data. The ultimate goal is to develop effective strategies for detecting and mitigating this pervasive fraud.

Research axis of the MLSec Learning Network Systems Security (MLNS2) INRIA Associate Team

Domain-wide recognized by their high value in human presence and activity studies, cellular network datasets (i.e., eXtended Data Records or XDRs), however, present accessibility, usability, and privacy issues. These hurdles hinder their effective utilization and impede research reproducibility. This project explores AI-powered strategies to make available to the community realistic xDRs being complete, i.e., including both traffic and mobility information. Notably, achieving the desired fine-grained granularity through per-individual data description poses a privacy challenge, which we thoroughly evaluate.

Research axis of the STIC-AmSud LINT Associate Team