Explainable machine learning attracts increasing attention as it improves transparency of models, which is helpful for machine learning to be trusted in real applications. However, explanation methods have recently been demonstrated to be vulnerable to manipulation, where we can easily change a model's explanation while keeping its prediction constant. To tackle this problem, some efforts have been paid to use more stable explanation methods or to change model configurations. In this work, we tackle the problem from the training perspective, and propose a new training scheme called Adversarial Training on EXplanations (ATEX) to improve the internal explanation stability of a model regardless of the specific explanation method being applied. Instead of directly specifying explanation values over data instances, ATEX only puts requirement on model predictions which avoids involving second-order derivatives in optimization. As a further discussion, we also find that explanation stability is closely related to another property of the model, i.e., the risk of being exposed to adversarial attack. Through experiments, besides showing that ATEX improves model robustness against manipulation targeting explanation, it also brings additional benefits including smoothing explanations and improving the efficacy of adversarial training if applied to the model. The codes of our work are available at: https://github.com/DefIntpMan.

In this paper, we introduce DSN (Deep Serial Number), a new watermarking approach that can prevent the stolen model from being deployed by unauthorized parties. Recently, watermarking in DNNs has emerged as a new research direction for owners to claim ownership of DNN models. However, the verification schemes of existing watermarking approaches are vulnerable to various watermark attacks. Different from existing work that embeds identification information into DNNs, we explore a new DNN Intellectual Property Protection mechanism that can prevent adversaries from deploying the stolen deep neural networks. Motivated by the success of serial number in protecting conventional software IP, we introduce the first attempt to embed a serial number into DNNs. Specifically, the proposed DSN is implemented in the knowledge distillation framework, where a private teacher DNN is first trained, then its knowledge is distilled and transferred to a series of customized student DNNs. During the distillation process, each customer DNN is augmented with a unique serial number, i.e., an encrypted 0/1 bit trigger pattern. Customer DNN works properly only when a potential customer enters the valid serial number. The embedded serial number could be used as a strong watermark for ownership verification. Experiments on various applications indicate that DSN is effective in terms of preventing unauthorized application while not sacrificing the original DNN performance. The experimental analysis further shows that DSN is resistant to different categories of attacks.