Random and Adversarial Bit Error Robustness: Energy-Efficient and Secure DNN Accelerators.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Hardware Security

Background:

• Deep neural network (DNN) accelerators offer energy savings but face challenges with low-voltage operation leading to bit failures.
• These accelerators are susceptible to adversarial attacks targeting voltage controllers or individual bits, compromising their integrity.
• Existing solutions often require hardware modifications or lack generalizability across different operating conditions.

Purpose of the Study:

• To develop a robust training methodology for DNN accelerators that enhances resilience against bit errors and adversarial attacks.
• To achieve significant energy savings through low-voltage operation and low-precision quantization without compromising accuracy.
• To improve the overall security and reliability of DNN accelerators in diverse operational environments.

Main Methods:

• Implementing robust fixed-point quantization and weight clipping techniques.
• Employing random bit error training (RandBET) or adversarial bit error training (AdvBET) for quantized DNN weights.
• Developing a novel adversarial bit error attack to test and validate robustness.

Main Results:

• Achieved significant energy reductions of 20%/30% for 8/4-bit quantization on CIFAR10 with minimal accuracy loss (0.8%/2%).
• Demonstrated robustness against both random and adversarial bit errors, significantly improving security.
• Reduced test error from over 90% to 26.22% even with up to 320 adversarial bit errors.

Conclusions:

• The proposed training methods (RandBET/AdvBET) effectively enhance DNN accelerator robustness against bit errors and adversarial attacks.
• The approach offers substantial energy savings for low-voltage and low-precision operation while improving security.
• This generalized solution does not require hardware changes and outperforms related work in terms of resilience and applicability.