Semantic Correlation Descriptors Identify Training Datasets

A new white-box fingerprinting method called semantic correlation descriptors (SCDs) can identify which dataset a model was trained on by analyzing the spurious correlations it internalizes. Researchers argue that datasets leave unique traces in a model's learned semantic correlation structure—incidental regularities predictive within a dataset but not causal for the underlying task. This approach moves beyond existing dataset-level membership inference methods that rely on confidence scores, losses, margins, generated samples, or query responses. In controlled leave-one-dataset-out diagnostics, SCDs perfectly separate matching from non-matching dataset pairs.