Encrypted Image Filtering Using Homomorphic Encryption

Zama has developed a new Hugging Face space utilizing Concrete-Numpy and Concrete-ML to apply homomorphic filters over images, ensuring data remains encrypted during transit and processing. Concrete-Numpy enables computations on encrypted data without decryption, while Concrete-ML allows the application of machine learning models in a Fully Homomorphic Encryption (FHE) setting without requiring cryptographic expertise. Users can create image processing filters, such as sharpening or noise reduction, using Torch models and compile them into FHE circuits. These filters can be deployed via a Client-Server interface, maintaining privacy by encrypting images before processing, with both the client and server components capable of handling encrypted data. A demo on Hugging Face showcases this functionality, where images are encrypted, processed, and decrypted, with the server having no access to the original or processed images. This approach ensures privacy and security, making it suitable for sensitive image processing applications.