Bottlenecks in Secure Adoption of Deep Neural Networks in Safety-Critical Applications

The widespread proliferation of deep learning models has resulted in their extensive adoption in various domains, ranging from image recognition, Natural Language Processing (NLP), medical diagnosis, autonomous driving etc. Features such as high accuracy, quick inference, high scalability and adaptability have been instrumental in this revolution. Nonetheless, there are several elements that pose a threat to secure and reliable implementation of these models in safety-critical applications. Three primary reasons can be attributed to engender this bottleneck. First, Deep Neural Networks (DNN) are vulnerable to attacks that manipulate inputs and/or network parameters, which can compromise their performance. Additionally, faults in the DNN hardware can have adverse effects on the network accuracy. Next, error and uncertainty quantification in DNN-based inference are complex, which can undermine the dependability of their decisions. Finally, the lack of transparency and explainability in black-box models impedes their adoption in highly critical environments. This paper offers a comprehensive exploration of these issues, which must be addressed for improving the safety and reliability in adoption of machine learning models in mission-critical applications.