Machine learning for embedded devices software analysis via hardware platform emulation

Nowadays commercial-off-the-shelf (COTS) embedded devices are widely used in many security-critical systems like nuclear stations and traffic control systems. Most of these devices has proprietary hardware and software (frequently called firmware) with little documentation available. Another common feature is the use of “binary blob” firmware, where hardware specific and high level layers can't be easily separated and, therefore are forced to be analyzed together. All these facts make firmware analysis quite a challenging task. In this paper, we'll suggest an approach for the firmware analysis with poorly documented hardware platform emulation. The advantages of this approach are almost full control under firmware state, achieving easy to scale fuzzing and manual bug hunting facilitation. For the purpose of successful realization, an identification of communication between hardware components (e.g. communication between main SoC and Bluetooth SoC) should be done. To address this issue, we suggest the use of machine learning, which, because of its nature, enables construction of algorithms that can learn from and make predictions on data.