NeuroUnlock: Unlocking the Architecture of Obfuscated Deep Neural Networks

TL;DR

NeuroUnlock reveals vulnerabilities in current DNN obfuscation methods by effectively reversing them to recover original architectures, which enhances the success of subsequent attacks and highlights the need for more resilient obfuscation techniques.

Contribution

This work introduces NeuroUnlock, a novel sequence-to-sequence attack that can reverse state-of-the-art DNN obfuscation, and proposes ReDLock, a more resilient obfuscation method.

Findings

NeuroUnlock successfully recovers architectures of various obfuscated DNNs.

ReDLock increases resilience, reducing NeuroUnlock's success rate.

Recovered models maintain high accuracy with minimal performance loss.

Abstract

The advancements of deep neural networks (DNNs) have led to their deployment in diverse settings, including safety and security-critical applications. As a result, the characteristics of these models have become sensitive intellectual properties that require protection from malicious users. Extracting the architecture of a DNN through leaky side-channels (e.g., memory access) allows adversaries to (i) clone the model, and (ii) craft adversarial attacks. DNN obfuscation thwarts side-channel-based architecture stealing (SCAS) attacks by altering the run-time traces of a given DNN while preserving its functionality. In this work, we expose the vulnerability of state-of-the-art DNN obfuscation methods to these attacks. We present NeuroUnlock, a novel SCAS attack against obfuscated DNNs. Our NeuroUnlock employs a sequence-to-sequence model that learns the obfuscation procedure and automatically reverts it, thereby recovering the original DNN architecture. We demonstrate the effectiveness of NeuroUnlock by recovering the architecture of 200 randomly generated and obfuscated DNNs running on the Nvidia RTX 2080 TI graphics processing unit (GPU). Moreover, NeuroUnlock recovers the architecture of various other obfuscated DNNs, such as the VGG-11, VGG-13, ResNet-20, and ResNet-32 networks. After recovering the architecture, NeuroUnlock automatically builds a near-equivalent DNN with only a 1.4% drop in the testing accuracy. We further show that launching a subsequent adversarial attack on the recovered DNNs boosts the success rate of the adversarial attack by 51.7% in average compared to launching it on the obfuscated versions. Additionally, we propose a novel methodology for DNN obfuscation, ReDLock, which eradicates the deterministic nature of the obfuscation and achieves 2.16X more resilience to the NeuroUnlock attack. We release the NeuroUnlock and the ReDLock as open-source frameworks.