Extracting the architecture of layers of a given deep neural network (DNN) through hardware-based side channels allows adversaries to steal its intellectual property and even launch powerful adversarial attacks on the target system. In this work, we propose DNN-Alias, an obfuscation method for DNNs that forces all the layers in a given network to have similar execution traces, preventing attack models from differentiating between the layers. Towards this, DNN-Alias performs various layer-obfuscation operations, e.g., layer branching, layer deepening, etc, to alter the run-time traces while maintaining the functionality. DNN-Alias deploys an evolutionary algorithm to find the best combination of obfuscation operations in terms of maximizing the security level while maintaining a user-provided latency overhead budget. We demonstrate the effectiveness of our DNN-Alias technique by obfuscating the architecture of 700 randomly generated and obfuscated DNNs running on multiple Nvidia RTX 2080 TI GPU-based machines. Our experiments show that state-of-the-art side-channel architecture stealing attacks cannot extract the original DNN accurately. Moreover, we obfuscate the architecture of various DNNs, such as the VGG-11, VGG-13, ResNet-20, and ResNet-32 networks. Training the DNNs using the standard CIFAR10 dataset, we show that our DNN-Alias maintains the functionality of the original DNNs by preserving the original inference accuracy. Further, the experiments highlight that adversarial attack on obfuscated DNNs is unsuccessful.
View on arXiv