Jerome Lenssen, Alexandru Paler (Apr 29 2025).

Abstract: Neural network decoders are becoming essential for achieving fault-tolerant quantum computations. However, their internal mechanisms are poorly understood, hindering our ability to ensure their reliability and security against adversarial attacks. Leading machine learning decoders utilize recurrent and transformer models (e.g., AlphaQubit), with reinforcement learning (RL) playing a key role in training advanced transformer models (e.g., DeepSeek R1). In this work, we target a basic RL surface code decoder (DeepQ) to create the first adversarial attack on quantum error correction. By applying state-of-the-art white-box methods, we uncover vulnerabilities in this decoder, demonstrating an attack that reduces the logical qubit lifetime in memory experiments by up to five orders of magnitude. We validate that this attack exploits a genuine weakness, as the decoder exhibits robustness against noise fluctuations, is largely unaffected by substituting the referee decoder, responsible for episode termination, with an MWPM decoder, and demonstrates fault tolerance at checkable code distances. This attack highlights the susceptibility of machine learning-based QEC and underscores the importance of further research into robust QEC methods.

Arxiv: https://arxiv.org/abs/2504.19651