Deep Learning Gradient Visualization-Based Pre-Silicon Side-Channel Leakage Location

While side-channel attacks (SCAs) have become a significant threat to cryptographic algorithms, masking is considered as an effective countermeasure against SCAs. On the one hand, securely implementing the scheme is a challenging and error-prone task. It is essential to detect leakage in a complicated cryptographic circuit. However, the traditional method of leakage detection is always inaccuracy or time consumption. On the other hand, the deep learning-based power attacks have shown their threat to the masking without combining functions. Compared to the leakage detection done under the traditional provable security framework, the security evaluation against deep learning-based attacks at the pre-silicon stage has not been discussed. To this end, this paper investigates the strategies of leveraging the deep learning techniques to achieve an efficient leakage location method. In this paper, we present the first approach utilizing deep learning-based leakage location for both unprotected and protected implementations at the pre-silicon stage. Firstly, we propose the leakage location method named Gradient Visualization-based location (GVL), which provides leakage location at the different levels of design. Gradient visualization is known as a sensitivity analysis method to understand better how a natural network can learn to predict the sensitive label based on the input. We theoretically show how the gradient visualization can be used to locate leakage components in the netlist efficiently. Moreover, we link the result with the metric in deep learning-based leakage assessment, which fills the lack of leakage evaluation at the pre-silicon stage against deep learning-based SCAs. We further confirm the effectiveness of the proposed method on unprotected implementation, low entropy masked implementation, and provable secure masked implementation. The results show that the proposed methodology outperforms the traditional location methods in the masked cases, where the time consumption is reduced by about 2x to 10x with fewer false negatives and no false positives.