LiBRA: Detection-Aware Image Watermark Removal via Bidirectional Latent Optimization
LiBRA attack removes AI-image watermarks in an autoencoder's latent space while preserving quality, steering decoder confidence toward random guessing instead of inversion.