Advances in Deep Learning Security and Watermarking

The field of deep learning security and watermarking is rapidly evolving, with a focus on developing robust and secure methods for protecting intellectual property and detecting malicious activities. Recent research has explored innovative approaches to neural network watermarking, including the use of frequency components and cryptographic chains to create secure and robust watermarks. Additionally, there have been significant advancements in deepfake detection, with new methods leveraging coarse-to-fine spatial information, semantic information, and feature orthogonality to improve generalization and detection capabilities. Noteworthy papers in this area include those proposing novel watermarking schemes, such as ChainMarks, and those introducing new deepfake detection strategies, such as Cross-Branch Orthogonality and RealID. These developments have the potential to significantly impact the field, enabling more effective protection of deep learning models and detection of malicious activities.

Sources

Towards the Resistance of Neural Network Watermarking to Fine-tuning

Watermark Overwriting Attack on StegaStamp algorithm

MGFF-TDNN: A Multi-Granularity Feature Fusion TDNN Model with Depth-Wise Separable Module for Speaker Verification

MoDE: Mixture of Diffusion Experts for Any Occluded Face Recognition

DATA: Multi-Disentanglement based Contrastive Learning for Open-World Semi-Supervised Deepfake Attribution

Learning Real Facial Concepts for Independent Deepfake Detection

Cross-Branch Orthogonality for Improved Generalization in Face Deepfake Detection

ChainMarks: Securing DNN Watermark with Cryptographic Chain

SSH-Net: A Self-Supervised and Hybrid Network for Noisy Image Watermark Removal

Built with on top of