Secure CNN Accelerator

Deep learning techniques like Convolution Neural Networks (CNN) are nowadays very popular in applications like object detection, semantic segmentation etc. The differentiation among various CNN based solutions come mainly from training data, which gets translated to CNN model parameters. The traditional approach to secure the model by decoupling cryptography from inference has inherent disadvantages of higher boot time, possible DRAM snooping attack and complexity in working with multiple vendors on given SoC. This paper presents novel architecture to overcome these limitations by just-in-time decryption. The proposed solution uses novel techniques namely encryption of intermediate tensors, dynamic block chaining of tensors & weight and hardware key management. The proposed solution is designed and simulated for 28 nm silicon process node running at 600 MHz with synthesis area of ~1 mm2 confirming removal of mentioned limitations of traditional solutions.