An Efficiency Evaluation of Far-Field Electromagnetic Deep Learning Side-Channel Attacks in Controlled Environments

Abstract: As more and more modern systems and products use built-in microcontrollers, hardware security becomes more important to protect against cyber-attacks. Internet of things devices, like Bluetooth devices, usually use an encryption algorithm to keep data safe from hackers. Advanced Encryption Standard (AES) is a commonly used encryption algorithm. AES itself is hard to break. However, it is possible to utilize the information leaking from a system during the execution of encryption, called side-channel, to recover the key or part of the key used by the encryption algorithm. This kind of attack is called a side-channel attack (SCA). In this study, two deep learning (DL) models are trained to attack the Bluetooth microcontroller unit Nordic nRF52 development kit equipped with an nRF52832 chip. The DL models are trained using the far-field electromagnetic emissions that the microcontroller unintentionally generates and transmits through the antenna while encrypting data. All encryptions are executed with a fixed key and random plaintext. The attack is conducted in two stages: the profiling and attack stages. In the profiling stage, where the attacker is assumed to have full system control, 100 000 traces holding encryption information are sampled and used to train the DL models to classify a sub-byte of the fixed key given a trace. In the attack stage, traces are captured in two different environments. The first is an entirely isolated environment, while the second adds a specific Wi-Fi access point and client connection that execute HTTP requests and responses in this isolated environment referred to as the system environment. Given traces obtained from one of the two attack environments, the performance of the trained models at classifying the correct sub-key is evaluated.  To summarize the results of this study, twelve SCAs are performed on six datasets captured in two different environments using two different DL models for each dataset. The correct key byte can be retrieved in three of these SCAs. All three successful attacks are made in an isolated environment without any interfering noise. The best performance is achieved with the multi-layer perceptron DL architecture, processing traces each composed of 10 averaged traces of the identical encryption, and the correct key-byte is recovered after 8198 traces.