A team of researchers has demonstrated that it is possible to steal an
artificial intelligence (AI) model without actually gaining access to the device that is running the model. The uniqueness of the technique lies in the fact that it works efficiently even if the thief may not have any prior knowledge as to how the AI works in the first place, or how the computer is structured.
According to North Carolina State University's Department of Electrical and Computer Engineering, the method is known as TPUXtract, and it is a product of their department. With the help of a team of four scientists, who used high-end equipment and a technique known as "online template-building", they were able to deduce the hyperparameters of a convolutional neural network (CNN) running on Google Edge Tensor Processing Unit (TPU), which is the settings that define its structure and behaviour, with a 99.91% accuracy rate.
The TPUXtract is an advanced side-channel attack technique devised by researchers at the North Carolina State University, designed to protect servers from attacks. A convolutional neural network (CNN) running on a Google Edge Tensor Processing Unit (TPU) is targeted in the attack, and electromagnetic signals are exploited to extract hyperparameters and configurations of the model without the need for previous knowledge of its architecture and software.
A significant risk to the security of AI models and the integrity of intellectual property is posed by these types of attacks, which manifest themselves across three distinct phases, each of which is based on advanced methods to compromise the AI models' integrity. Attackers in the Profiling Phase observe and capture side-channel emissions produced by the target TPU as it processes known input data as part of the Profiling Phase. As a result, they have been able to decode unique patterns which correspond to specific operations such as convolutional layers and activation functions by using advanced methods like Differential Power Analysis (DPA) and Cache Timing Analysis.
The Reconstruction Phase begins with the extraction and analysis of these patterns, and they are meticulously matched to known processing behaviours This enables adversaries to make an inference about the architecture of the AI model, including the layers that have been configured, the connections made, and the parameters that are relevant such as weight and bias.
Through a series of repeated simulations and output comparisons, they can refine their understanding of the model in a way that enables precise reconstruction of the original model.
Finally, the Validation Phase ensures that the replicated model is accurate.
During the testing process, it is subject to rigorous testing with fresh inputs to ensure that it performs similarly to that of the original, thus providing reliable proof of its success.
The threat that TPUXtract poses to intellectual property (IP) is composed of the fact that it enables attackers to steal and duplicate artificial intelligence models, bypassing the significant resources that are needed to develop them.
The competition could recreate and mimic models such as ChatGPT without having to invest in costly infrastructure or train their employees.
In addition to IP theft, TPUXtract exposed cybersecurity risks by revealing an AI model's structure that provided visibility into its development and capabilities. This information could be used to identify vulnerabilities and enable cyberattacks, as well as expose sensitive data from a variety of industries, including healthcare and automotive.
Further, the attack requires specific equipment, such as a Riscure Electromagnetic Probe Station, high-sensitivity probes, and Picoscope oscilloscope, so only well-funded groups, for example, corporate competitors or state-sponsored actors, can execute it. As a result of the technical and financial requirements for the attack, it can only be executed by well-funded groups.
With the understanding that any electronic device will emit electromagnetic radiation as a byproduct of its operations, the nature and composition of that radiation will be affected by what the device does.
To conduct their experiments, the researchers placed an EM probe on top of the TPU after removing any obstructions such as cooling fans and centring it over the part of the chip emitting the strongest electromagnetic signals. The machine then emitted signals as a result of input data, and the signals were recorded.
The researchers used the Google Edge TPU for this demonstration because it is a commercially available chip that is widely used to run AI models on edge devices meaning devices utilized by end users in the field, as opposed to AI systems that are used for database applications.
During the demonstration, electromagnetic signals were monitored as a part of the technique used to conduct the demonstration.
A TPU chip was placed on top of a probe that was used by researchers to determine the structure and layer details of an AI model by recording changes in the electromagnetic field of the TPU during AI processing. The probe provided real-time data about changes in the electromagnetic field of the TPU during AI processing. To verify the model's electromagnetic signature, the researchers compared it to other signatures made by AI models made on a similar device - in this case, another Google Edge TPU.
Using this technique, Kurian says, AI models can be stolen from a variety of different devices, including smartphones, tablets and computers.
The attacker should be able to use this technique as long as they know the device from which they want to steal, have access to it while it is running an AI model, and have access to another device with similar specifications According to Kurian, the electromagnetic data from the sensor is essentially a ‘signature’ of the way AI processes information.
There is a lot of work that goes into pulling off TPUXtract. The process not only requires a great deal of technical expertise, but it also requires a great deal of expensive and niche equipment as well. To scan the chip's surface, NCSU researchers used a Riscure EM probe station equipped with a motorized XYZ table, and a high-sensitivity electromagnetic probe to capture the weak signals emanating from it.
It is said that the traces were recorded using a Picoscope 6000E oscilloscope, and Riscure's icWaves FPGA device aligned them in real-time, and the icWaves transceiver translated and filtered out the irrelevant signals using bandpass filters and AM/FM demodulation, respectively. While this may seem difficult and costly for a hacker to do on their own, Kurian explains, "It is possible for a rival company to do this within a couple of days, regardless of how difficult and expensive it will be.
Taking the
threat of TPUXtract into account, this model poses a formidable challenge to AI model security, highlighting the importance of proactive measures. As an organization, it is crucial to understand how such attacks work, implement robust defences, and ensure that they can safeguard their intellectual property while maintaining trust in their artificial intelligence systems. The AI and cybersecurity communities must learn continuously and collaborate to stay ahead of the changing threats as they arise.
