← Latest papers
⚛️ quantum physics

Evaluating Quantum Circuit Protection Methods After Compilation: An Empirical Comparison

This paper empirically demonstrates that evaluating quantum circuit protection methods only before compilation is insufficient, as most existing schemes lose their structural disguise or fail to hide outputs after optimization, whereas the proposed Q-MaskCut technique successfully maintains output privacy and structural obfuscation even after compilation and provider recompilation.

Original author: Ing Muyleang, Leanghok Hour, Youngsun Han

Published 2026-11-13
📖 6 min read🧠 Deep dive

Original author: Ing Muyleang, Leanghok Hour, Youngsun Han

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the emerging world of quantum computing, powerful machines are beginning to solve problems that would take classical computers thousands of years. Because these machines are expensive and delicate, most people access them not by owning one, but by sending instructions over the internet to a cloud provider. A user writes a set of instructions, known as a circuit, which tells the machine how to manipulate tiny units of information called qubits. The provider runs the circuit and sends back the results. However, this arrangement creates a privacy gap: the provider sees the entire set of instructions and the final answer. If the instructions contain a secret algorithm or the answer holds confidential data, the user has no way to hide them from the company operating the machine.

Researchers have long tried to solve this by adding "decoys" or extra steps to the instructions that cancel each other out, hoping to confuse anyone watching. The problem is that before the machine runs, a sophisticated software program rewrites the instructions to make them run faster and more efficiently. This process, called compilation, often spots the extra steps and removes them, leaving the secret exposed. A new study published in Physica Scripta by Ing Muyleang, Leanghok Hour, and Youngsun Han investigates whether these protection methods actually survive this rewriting process. They tested several existing techniques and a new combination of methods to see if they could truly hide a user's work from a cloud provider who might be curious about what is being calculated.

The researchers focused on a specific goal: keeping the provider from learning the user's algorithm and the answer, without trying to stop a malicious attack. They treated the provider as "honest but curious," meaning the company would run the job exactly as asked and return the true results, but would also look closely at everything it received to figure out what the user was doing. To test this, the team took a variety of quantum circuits, ranging from very small ones with just three qubits to large ones with up to 127 qubits. They applied different protection schemes to these circuits and then forced them through the same high-level optimization software that a real provider would use. Only after this rewriting did they measure how well the secrets were hidden and whether the final answer remained correct.

The study found that many popular protection methods fail completely once the software rewrites the code. Techniques that rely on inserting extra gates that cancel each other out, such as "composite" or "delayed" gates, were almost entirely stripped away by the optimizer. The resulting circuits looked nearly identical to the original, unhidden versions, leaving the provider able to read the algorithm and the answer. Another method, which tried to hide the structure by making the circuit much longer, did survive the rewriting but made the job take so long to run—up to 112 times longer—that it became impractical for current devices. A different approach called ECQCO, which combines encryption with structural changes, managed to hide the structure somewhat, but the researchers found that the code provided by its original authors could not successfully decode the final answer, rendering the protection useless for the user.

The team then evaluated a new combination of techniques they called Q-MaskCut. This method does not rely on a single trick but layers several existing tools together. First, it adds a secret key that flips the final answer bits, so the provider sees a scrambled result. It also inserts "decoy" wires that look like real data wires but carry no information, making it hard to tell which parts of the circuit are doing the actual work. The circuit is then cut into smaller pieces that are run separately, and after the software rewrites them, the researchers add a layer of masking pulses that fill the empty time between operations. These pulses are designed to look like noise but actually hide the structure of the circuit.

The results showed that Q-MaskCut was the only method that successfully hid both the structure and the answer while still allowing the user to recover the correct result. When the provider looked at the scrambled output without the secret key, the results looked completely random, with no hint of the true answer. When the user applied their key, they recovered the exact correct answer. In simulations without noise, the method hid the output so well that the scrambled results were almost entirely unrelated to the true answer. Even when the researchers simulated the noise found in real machines, the method performed just as well as the unprotected version, meaning the user did not lose accuracy in exchange for privacy.

However, the study also revealed a critical limitation. The structural disguise provided by Q-MaskCut relies heavily on the masking pulses added after the software rewrites the code. If the provider were to take the submitted circuit and run it through the optimization software a second time, these masking pulses would be removed, and the structural disguise would drop significantly. While the secret key and decoy wires would still protect the answer, the visual appearance of the circuit would become much more similar to the original. This suggests that while the method works against a provider who simply runs the job as submitted, it is not a permanent shield against a provider who actively re-optimizes the code.

The researchers concluded that to truly evaluate how well a protection method works, it must be tested after the circuit has been rewritten by the provider's software, and again if the provider rewrites it a second time. They found that many methods that look strong on paper disappear once the optimization software does its work. Q-MaskCut emerged as the most effective solution tested, capable of hiding the answer and the structure without slowing down the single execution time, though it requires the user to run the job many more times—between seven and ninety-one times depending on the circuit size—and to perform a complex calculation on their own computer to piece the results back together. The study serves as a reality check for the field, showing that protecting quantum computations in the cloud is possible, but only if the protection is designed to survive the very software that makes quantum computers useful in the first place.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →