BlackCATT: Black-Box Collusion-Aware Traitor Tracing in Federated Learning

The paper proposes BlackCATT, a black-box collusion-aware traitor tracing method for federated learning, which the authors claim is the first general collusion-resistant embedding scheme. It introduces a collusion-aware embedding loss term and iteratively optimizes triggers rather than using a fixed trigger set, to improve convergence and tracing performance. Experiments confirm its effectiveness across different architectures and datasets; BlackCATT+FR performs function regularization via auxiliary samples at the aggregation end, mitigating incompatibility in main-task updates after models containing batch normalization layers learn different watermarks, while maintaining tracing performance.