Fingerprinting Models

Model Fingerprinting is a technique for identifying and verifying a model's identity based on its unique behavioral characteristics — detecting whether a suspect model is a copy, derivative, or extraction of a protected model by probing its behavior on specially designed inputs.

Fingerprinting Methods

• Conferrable Examples: Find inputs where the original model and its derivatives agree but other models disagree.
• Decision Boundary Analysis: Probe the model's decision boundaries — stolen models have similar boundary geometry.
• Adversarial Examples: Adversarial examples that transfer from the original model to its copies can serve as fingerprints.
• Statistical Tests: Compare confidence distributions, error patterns, or calibration curves.

Why It Matters

• No Cooperation: Unlike watermarking (which requires embedding during training), fingerprinting works post-hoc.
• Copy Detection: Identify model theft even when the stolen model has been fine-tuned or distilled.
• Legal Evidence: Provide forensic evidence of model copying for intellectual property disputes.

Model Fingerprinting is behavioral identification — recognizing a model's unique "personality" to detect copies without requiring embedded watermarks.