Deep learning in many aspects of the Internet of Things

Deep learning has become a way to solve many challenging problems. In terms of target detection, speech recognition and language translation, deep learning is by far the best performing method. Many people regard deep neural networks (DNNs) as magical black boxes. We enter some data and come out with our solutions! In fact, things are much more complicated.

1) Image recognition:

In a large part of the application scenario of IoT, the input deep learning data is a picture or video. Every day, everyone uses their high-definition camera to capture pictures and videos from their mobile phones. In addition, smart cameras are used in homes, campuses or factories. Therefore, image recognition, classification, and target detection are the basic applications of such devices.

2) Speech recognition

With the popularity of smartphones and wearables, speech recognition has become a natural and convenient way for people to interact with their devices. Price et al. built a dedicated low-power deep learning chip for automatic speech recognition. The energy consumption of this special chip is 100 times lower than that of the speech recognition tool currently running on mobile phones.

3) Indoor positioning

Indoor positioning has many applications in the IoT field, such as smart homes, smart campuses, or smart hospitals. For example, the DeepFi system, in the online training phase, uses the previously stored WiFi channel state information data to train the network weight through deep learning, and the user positioning is determined by fingerpringTIng in the online positioning phase.

4) Physiological and psychological state detection

The combination of IoT and deep learning is also used to detect various physical or mental states such as posture, activity and mood. Many IoT applications integrate human pose estimation or activity recognition modules into the delivered services, such as smart homes, smart cars, XBox, health, sports and more.

5) Security and privacy

Security and privacy are an important concern for all IoT applications. In fact, the effectiveness of system functionality depends on whether it protects machine learning tools and processes from attack. False Data InjecTIon (FDI) is a common type of attack in data-driven systems. He et al. proposed using conditional DBN to extract FDI features from historical data and then use these features for real-time attack detection. As a major contributor to IoT data and applications, smartphones are also threatened by hackers. Yuan et al. proposed to use a deep learning framework to identify malware in Android applications with an accuracy rate of 96.5%. The security and privacy protection of deep machine learning methods is the most important factor in the application of IoT. Shokri et al. proposed a method to solve the privacy protection problem of deep learning model for distributed learning.