Watermarking for Data Provenance in Object Detection

Machine Learning (ML) has pervaded almost every industry including banking, autonomous vehicles, healthcare, retail, and defense. Effective use of ML can be a huge competitive advantage, and performance of an ML model is dependent on massive, well-constructed, annotated datasets used for training. These datasets are often sensitive and contain proprietary data. They can be costly to collect and require subject matter experts to annotate. Thus, annotated datasets are regarded as highly valuable intellectual property that must be protected. The owners of a proprietary training dataset need the ability to track its use in the wild to safeguard their intellectual property. We propose a novel methodology for object detection dataset traceability through ‘watermarking’ the training data. This watermark is embedded into object instances within a training dataset, which the model then learns. We design the watermark such that the model is sensitive to its presence within an image, even after the model is fine-tuned on another task. Additionally, the watermark is designed for negligible model performance fluctuations when compared to identical models trained on non-watermarked data. We demonstrate the effectiveness of our technique on multiple object detection model architectures, and conclude that this methodology can provide a means to effectively track dataset usage.