TL;DR
Brain-Computer Interfaces (BCIs) that allow users to type with their thoughts, particularly P300 and SSVEP spellers, are surprisingly vulnerable to tiny, imperceptible "adversarial perturbations." These subtle manipulations can trick the BCI into spelling anything an attacker desires, posing a significant security risk with potentially serious consequences for users.
What Problem Does This Paper Address?
Brain-Computer Interfaces (BCIs) offer a lifeline for individuals with severe disabilities, enabling them to communicate and interact with the world through their brain signals. While much research has focused on making these BCIs faster and more accurate, their security has largely been overlooked. This paper highlights a critical vulnerability: the susceptibility of BCI spellers to adversarial attacks. These attacks involve introducing minuscule, human-imperceptible noise into the brain signals that can deliberately mislead the BCI into producing incorrect outputs, ranging from user frustration to potentially severe misinterpretations in clinical settings.
Methods
The researchers developed a novel method for generating "adversarial perturbation templates" for both P300 and SSVEP BCI spellers. Unlike previous adversarial attack methods that often require knowing the entire signal or are not causal, their approach focuses on creating a fixed perturbation template that can be applied to incoming EEG data, ensuring causality and practical implementation.
For P300 spellers, they identified directions pointing from non-target EEG epochs towards the decision boundary of the BCI's classifier. These directions were summed to create a universal perturbation. For SSVEP spellers, which rely on identifying specific flickering frequencies, the challenge was greater due to multiple classes and varying classifier parameters. They employed a method to maximize the correlation between the perturbed EEG signal and reference signals at the attacker's desired frequency.
Datasets
- P300 Speller Dataset: Publicly available dataset (Dataset II from Wolpaw et al.) with 64-channel EEG signals from two subjects.
- P300 Speller Dataset (ALS Patients): Publicly available dataset from BNCI Horizon 2020, featuring eight-channel EEG signals from eight ALS patients.
- SSVEP Speller Dataset: Publicly available benchmark dataset from Tsinghua University, with 64-channel EEG signals from 35 subjects.
Key Insights and Ideas
- BCIs are not inherently secure: This paper demonstrates for the first time that common BCI spellers are highly vulnerable to adversarial attacks, a concern previously under-explored.
- Tiny noise, big impact: The adversarial perturbations are so small they are virtually undetectable by humans, yet they can completely alter the BCI's output.
- Targeted attacks are feasible: Attackers can specifically design perturbations to make the BCI spell any desired character, not just introduce random errors.
- Causality matters: The proposed attack method is causal, meaning it can be applied in real-time as EEG data is acquired, making it a practical threat.
- SSVEP spellers are more transferable: Adversarial perturbations generated for SSVEP spellers showed better transferability across different subjects compared to P300 spellers.
Key Results
- P300 Spellers: Adversarial attacks achieved an average success rate of over 90% in misleading the speller to output a specific attacker character, significantly reducing user accuracy to near zero. The SPR (Signal-to-Perturbation Ratio) was high (over 20 dB), indicating the perturbations were imperceptible.
- SSVEP Spellers: For most subjects, adversarial attacks achieved 70-100% success rates in manipulating the speller's output. User accuracy was also drastically reduced. The SPR was comparable to random noise (around 25 dB).
- Cross-Subject Transferability: P300 speller attacks showed poor cross-subject transferability, while SSVEP speller attacks were more transferable.
- Cross-Model Transferability: Adversarial perturbations generated for one BCI pipeline could still effectively attack a different pipeline, demonstrating a broader threat.
Limitations and Future Work
- Subject/Model Specificity: Current attack methods require some subject- or model-specific EEG data to construct the perturbation template. Future work could explore transferability to overcome this.
- Synchronization Dependence: The P300 attack is sensitive to the precise timing of the perturbation relative to the stimulus. Future research could investigate attacks that are less sensitive to synchronization delays.
- Closed-Loop BCI Impact: The paper discusses the significant implications for closed-loop BCI applications, where constant misclassification could lead users to abandon the system or result in severe misdiagnosis for patients relying solely on BCIs.
- Broader BCI Applications: The authors suggest their findings are applicable beyond spellers to other BCI applications like neuro-ergonomics, cognitive state assessment, and control systems (e.g., wheelchair control), where errors could be critical.
- Defense Strategies: The authors emphasize that their goal is to highlight the security concern and encourage research into defense mechanisms against these adversarial attacks.
Summary updated