**7. Discussion**

We have chosen the Android operating system, since it is the most widely used operating system in the world, with more than 80% of the mobile market [38]. Moreover, the results presented in Figure 16 are valid on the date of the samples in the dataset. For a reminder, the collection of malware is registered in Omnidroid dates from 2012 to 2018. We have observed the relabeling impact clearly with more recent reports from VirusTotal on the evaluation metrics. Accordingly, we can wonder what will happen to the accuracy and the precision of neural networks in several years.

Moreover, the apps may have evolved too much to be run on an API 16 emulator, which is that of DroidBox [26], the tool for extracting dynamic features. API 16 (summer 2012) enables to run of relatively old apps for the period of summer 2020. However, it is possible that app developers consider this API too old and set the API minimum to 21. Thus, it would be impossible to run their apps on the emulator. However, this can be an advantage, since we can analyze old apps. In addition, the extraction of dynamic features can be difficult in the context of apps requiring identification (Facebook, Instagram, WhatsApp, Messenger type apps, etc.). In fact, the automatic feature extraction tool would be blocked at the displaying of identification, and would not be able to explore all the app's functionalities. As a reminder, the emulator has a feature called Monkey that enables us to randomly click on the screen to simulate user clicks. The extracted features would be either nonexistent or in too little representative quantity to be able to make a prediction. This is an intrinsic limitation of dynamic analysis.
