Radioactive data: tracing through training
This addresses the need for dataset provenance and copyright protection in machine learning, providing a robust detection method for researchers and practitioners.
The paper tackles the problem of detecting whether a specific image dataset was used to train a model by introducing 'radioactive data', which makes imperceptible changes to the dataset to leave an identifiable mark on trained models. Experiments on ImageNet with standard architectures show detection with high confidence (p<10^-4) even when only 1% of the training data is radioactive, offering a higher signal-to-noise ratio than existing methods.
We want to detect whether a particular image dataset has been used to train a model. We propose a new technique, \emph{radioactive data}, that makes imperceptible changes to this dataset such that any model trained on it will bear an identifiable mark. The mark is robust to strong variations such as different architectures or optimization methods. Given a trained model, our technique detects the use of radioactive data and provides a level of confidence (p-value). Our experiments on large-scale benchmarks (Imagenet), using standard architectures (Resnet-18, VGG-16, Densenet-121) and training procedures, show that we can detect usage of radioactive data with high confidence (p<10^-4) even when only 1% of the data used to trained our model is radioactive. Our method is robust to data augmentation and the stochasticity of deep network optimization. As a result, it offers a much higher signal-to-noise ratio than data poisoning and backdoor methods.