Board and Chip Diversity in Deep Learning Side-Channel Attacks : On ATtiny85 Implementations Featuring Encryption and Communication

Sammanfattning: Hardware security is an increasingly relevant topic because more and more systems and products are equipped with embedded microcontrollers. One type of threat against hardware security is attacks against encryption implementations in embedded hardware. The purpose of such attacks might be to extract the secret encryption key used to encrypt secret information that is being processed in the hardware. One type of such an attack that has gained more attention lately is side-channel attacks using deep learning algorithms. These attacks exploit the information that leaks from a chip in the form of the power the chip is consuming during encryption. In order to execute a side-channel attack assisted by deep learning, large amounts of data are needed for the neural network to train on. The data typically consists of several hundreds of thousands of power traces that have been captured from the profiling device. When the network has finished training, only a few power traces are required from a similar device to extract the key byte that has been used during encryption. In this project, the 8-bit microcontroller ATtiny85 was used as the victim device. AES-ECB 128 was used as the encryption algorithm. The goal of the project was to test how differences between boards and ATtiny85 chips affect the performance of side-channel attacks with deep learning. In the experiments, six different boards were used, where three of them had identical designs, and three of the boards had different designs. The data gathering was performed by measuring power consumption with an oscilloscope connected to a PC. The results showed that the similarity between the boards that were used for profiling and the boards that were attacked was the most important aspect for the attack to succeed with as few power traces as possible. If the board that was attacked was represented as a part of the training dataset, improved attack performance could be observed. If the training used data from several identical boards, no obvious improvement in attack performance could be seen. The results also showed that there are noticeable differences between identical ATtiny85 chips. These differences were obvious because the best attacks were the ones where the attacked chip was part of the training data set. There are several directions for future work, including how feasible these attacks are in real life scenarios and how to create efficient countermeasures.