Customized Watermarking for Deep Neural Networks via Label Distribution Perturbation

TL;DR

This paper introduces a novel DNN watermarking framework, USP and CSP, that embeds watermarks into models by perturbing output distributions, effectively resisting various attacks while maintaining high task performance.

Contribution

The paper proposes a new watermarking method that resists all known attacks, especially distillation, and balances watermark robustness with minimal impact on main task accuracy.

Findings

Resists all watermark removal attacks including distillation.

Achieves 98.68% watermark accuracy with only 0.59% impact on main task.

Outperforms existing methods in resisting distillation attacks.

Abstract

With the increasing application value of machine learning, the intellectual property (IP) rights of deep neural networks (DNN) are getting more and more attention. With our analysis, most of the existing DNN watermarking methods can resist fine-tuning and pruning attack, but distillation attack. To address these problem, we propose a new DNN watermarking framework, Unified Soft-label Perturbation (USP), having a detector paired with the model to be watermarked, and Customized Soft-label Perturbation (CSP), embedding watermark via adding perturbation into the model output probability distribution. Experimental results show that our methods can resist all watermark removal attacks and outperform in distillation attack. Besides, we also have an excellent trade-off between the main task and watermarking that achieving 98.68% watermark accuracy while only affecting the main task accuracy by 0.59%.