DINOHash: Learning Adversarially Robust Perceptual Hashes from Self-Supervised Features

Open-source software plays a vital role in machine learning research, yet projects focused on infrastructure and robustness often remain under-recognized. We present DINOHash, an open-source framework for robust perceptual image hashing designed for the provenance detection of AI-generated images. Unlike digital watermarking methods—which are easily broken by compression, cropping, or model leaks—DINOHash does not modify the image, making it inherently more secure, auditable, and resistant to transformation and adversarial attack. DINOHash is built on top of the open-sourced DINOv2 features and incorporates adversarial training to ensure resilience against adversaries. DINOHash offers a reliable foundation for researchers and practitioners to build secure content verification systems. To promote ease of adoption, the model is released in PyTorch and ONNX formats as well as on npm.