The explosion of data usage has contributed to the requirement of processing extensive amount of data for most of the applications on smart devices and edge- and fog computing nodes. Due to the scale and complexity of the tasks, decision support systems can greatly benefit from the use of machine learning (ML) techniques to correlate multimodal sensing to make accurate predictions and powerful inferences. Traditional ML algorithms have to be fed with previously extracted features. These features are usually identified in advance to reduce the complexity of the data and increase the visibility of the patterns to the learning algorithms. Furthermore, in some circumstances, like multiple object detection, the task needs to be divided into parts and solved individually and the partial results are combined at the final stage. The required human intervention and discontinuity in the process of accomplishing the tasks contribute to the reduced efficiency of the conventional ML algorithms in the face of massive raw data and intricate tasks. Deep learning (DL), also referred to as deep neural network (DNN), has overcome the weakness of the need for human's participation on effective feature identification and hard-core feature extraction. It learns the high-level features from raw data in an incremental manner and solves the problems end-to-end. As a result, DL has now become a preferred option for handling majority of the challenging tasks in image classification, speech recognition and language processing.

Chapter Contents:

• 9.1 Introduction
• 9.2 Preliminaries
• 9.2.1 Artificial neural networks (ANNs) and DNNs
• 9.2.2 Fundamental components of DNNs
• 9.2.3 Popular DNN architectures
• 9.2.4 Representative techniques for DNN hardware acceleration
• 9.3 Misprediction attacks
• 9.3.1 Threat model
• 9.3.2 Evasion attacks
• 9.3.2.1 Data evasion
• 9.3.2.2 Model evasion
• 9.3.3 Poisoning attacks
• 9.3.4 Backdoor attacks
• 9.4 Confidentiality attacks
• 9.4.1 Incentive
• 9.4.2 Model confidentiality attacks
• 9.4.3 Data confidentiality attacks
• 9.5 Explainability
• 9.5.1 Explainability of DNN processing
• 9.5.2 Explainability of DNN representations
• 9.5.3 Self-explainable systems
• 9.6 Conclusion
• Acknowledgment
• References

Inspec keywords: speech recognition; decision support systems; learning (artificial intelligence); image classification; computer network security; feature extraction; natural language processing

Other keywords: deep neural network; deep learning network security; image classification; high-level features; machine learning; speech recognition; smart devices; feature identification; hard-core feature extraction; multimodal sensing

Subjects: Computer communications; Image recognition; Natural language interfaces; Knowledge engineering techniques; Computer vision and image processing techniques; Data security; Computer networks and techniques; Decision support systems; Speech recognition and synthesis; Speech processing techniques