This paper proves that specific frequency components of convolutional filters are robust to fine-tuning and uses them as neural network fingerprinting.