New Attack Fingerprints Federated Learning Models via 5G Side Channels

This paper introduces FLINT, a novel black-box fingerprinting framework that can identify the architectural family of Federated Learning (FL) models, such as CNNs, RNNs, or Transformers. This is achieved by analyzing side-channel information from the 5G Physical (PHY) layer, specifically the scheduling metadata broadcast over the Physical Downlink Control Channel (PDCCH). Unlike previous attacks, FLINT does not require packet-level network visibility, which is often unavailable due to encryption and changing identifiers in 5G. FLINT overcomes these limitations by decoding PDCCH data, mapping dynamic Radio Network Temporary Identifiers (RNTIs) to specific user devices, and applying multi-view temporal modeling to discern architecture-specific training patterns. The ability to infer a client's model architecture from low-level 5G signals is a critical security vulnerability, as it can enable more targeted and effective downstream attacks. Experiments on a real 5G testbed demonstrated FLINT's high accuracy, achieving a macro F1-score of 0.930 for architecture classification.