DINAR: Enabling Distribution Agnostic Noise Injection in Machine Learning Hardware
Karthik Ganesan, Viktor Karyofyllis, Julianne Attai, Ahmed Hamoda, Natalie Enright Jerger · 2023
Machine learning (ML) has seen a major rise in popularity on edge devices in recent years, ranging from IoT devices to self-driving cars. Security in a critical consideration on these platforms. State-of-the-art security-centric ML algorithms (e.g., differentially private ML, adversarial robustness) require noise sampled from Laplace or Gaussian distributions. Edge accelerators lack CPUs [15, 25, 36, 50] to add such noise. Existing hardware approaches to generate noise on-the-fly incur high overheads and leak side-channel information that can undermine security [34, 47]. To remedy this, we propose DINAR,1 lightweight hardware that enables noise addition from arbitrary distributions. For differentially private ML, DINAR enables noise addition while incurring 23 × lower area and 40 × lower energy compared to producing noise directly on-chip.