Efficient On-Line Error Detection and Mitigation for Deep Neural Network Accelerators

• Published in: Computer Safety, Reliability, and Security pp. 205 - 219
• Main Authors: Schorn, Christoph, Guntoro, Andre, Ascheid, Gerd
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing
• Series: Lecture Notes in Computer Science
• ISBN: 9783319991290; 3319991299
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-99130-6_14

The use of deep neural network accelerators in safety-critical systems, for example autonomous vehicles, requires measures to ensure functional safety of the embedded hardware. However, due to the vast computational requirements that deep neural networks exhibit, the use of traditional redundancy-based approaches for the detection and mitigation of random hardware errors leads to very inefficient systems. In this paper we present an efficient and effective method to detect critical bit-flip errors in neural network accelerators and mitigate their effect at run time. Our method is based on an anomaly detection in the intermediate outputs of the neural network. We evaluate our method by performing fault injection simulations with two deep neural networks and data sets. In these experiments our error detector achieves a recall of up to 99.03% and a precision of up to 97.29%, while requiring a computation overhead of only 2.67% or less.