Search for: All records

Model stealing attacks on AI/ML devices undermine intellectual property rights, compromise the competitive advantage of the original model developers, and potentially expose sensitive data embedded in the model’s behavior to unauthorized parties. While previous research works have demonstrated successful side-channelbased model recovery in embedded microcontrollers and FPGA-based accelerators, the exploration of attacks on commercial ML accelerators remains largely unexplored. Moreover, prior side-channel attacks fail when they encounter previously unknown models. This paper demonstrates the first successful model extraction attack on the Google Edge Tensor Processing Unit (TPU), an off-the-shelf ML accelerator. Specifically, we show a hyperparameter stealing attack that can extract all layer configurations including the layer type, number of nodes, kernel/filter sizes, number of filters, strides, padding, and activation function. Most notably, our attack is the first comprehensive attack that can extract previously unseen models. This is achieved through an online template-building approach instead of a pre-trained ML-based approach used in prior works. Our results on a black-box Google Edge TPU evaluation show that, through obtained electromagnetic traces, our proposed framework can achieve 99.91% accuracy, making it the most accurate one to date. Our findings indicate that attackers can successfully extract various types of models on a black-box commercial TPU with utmost detail and call for countermeasures. 

Machine learning (ML) models can be trade secrets due to their development cost. Hence, they need protection against malicious forms of reverse engineering (e.g., in IP piracy). With a growing shift of ML to the edge devices, in part for performance and in part for privacy benefits, the models have become susceptible to the so-called physical side-channel attacks. ML being a relatively new target compared to cryptography poses the problem of side-channel analysis in a context that lacks published literature. The gap between the burgeoning edge-based ML devices and the research on adequate defenses to provide side-channel security for them thus motivates our study. Our work develops and combines different flavors of side-channel defenses for ML models in the hardware blocks. We propose and optimize the first defense based on Boolean masking . We first implement all the masked hardware blocks. We then present an adder optimization to reduce the area and latency overheads. Finally, we couple it with a shuffle-based defense. We quantify that the area-delay overhead of masking ranges from 5.4× to 4.7× depending on the adder topology used and demonstrate a first-order side-channel security of millions of power traces. Additionally, the shuffle countermeasure impedes a straightforward second-order attack on our first-order masked implementation. 

Intellectual Property (IP) thefts of trained machine learning (ML) models through side-channel attacks on inference engines are becoming a major threat. Indeed, several recent works have shown reverse engineering of the model internals using such attacks, but the research on building defenses is largely unexplored. There is a critical need to efficiently and securely transform those defenses from cryptography such as masking to ML frameworks. Existing works, however, revealed that a straightforward adaptation of such defenses either provides partial security or leads to high area overheads. To address those limitations, this work proposes a fundamentally new direction to construct neural networks that are inherently more compatible with masking. The key idea is to use modular arithmetic in neural networks and then efficiently realize masking, in either Boolean or arithmetic fashion, depending on the type of neural network layers. We demonstrate our approach on the edge-computing friendly binarized neural networks (BNN) and show how to modify the training and inference of such a network to work with modular arithmetic without sacrificing accuracy. We then design novel masking gadgets using Domain-Oriented Masking (DOM) to efficiently mask the unique operations of ML such as the activation function and the output layer classification, and we prove their security in the glitch-extended probing model. Finally, we implement fully masked neural networks on an FPGA, quantify that they can achieve a similar latency while reducing the FF and LUT costs over the state-of-the-art protected implementations by 34.2% and 42.6%, respectively, and demonstrate their first-order side-channel security with up to 1M traces.