Abstract
This paper investigates the security vulnerabilities of deep learning-based wireless receivers that employ online learning to adapt to dynamic, time-varying channels. While online adaptation is essential for maintaining performance in shifting environments, it introduces a critical attack vector: data poisoning. We propose a transfer-based black-box poisoning attack that requires no knowledge of the target receiver's internal architecture. By injecting subtle adversarial perturbations into pilot signals, an attacker can corrupt the receiver's learning process. Our evaluation across multiple channel models, including the high-fidelity COST 2100 model, demonstrates that these attacks significantly degrade communication reliability, with model-based receivers like DeepSIC and Meta-DeepSIC showing particular susceptibility.
Introduction: The Vulnerability of Online Adaptation
Modern wireless communication systems are increasingly adopting deep receivers—architectures that utilize Deep Neural Networks (DNNs) to perform signal detection and decoding. To remain effective in dynamic MIMO (Multiple-Input Multiple-Output) environments where channel conditions change rapidly, these receivers rely on online learning. During this process, the receiver uses a small set of known "pilot symbols" sent by the transmitter to update its internal parameters in real-time.
However, this reliance on external data for model updates creates a significant security risk. Most existing research on adversarial machine learning in wireless communications focuses on "evasion attacks," which attempt to fool a static model during the testing phase. This paper shifts the focus to "poisoning attacks," which target the training or update phase of the receiver.
As illustrated above, while evasion attacks (a) seek to cause misclassification of a specific input, poisoning attacks (b) aim to corrupt the model parameters ($\theta \rightarrow \theta^*$) so that the receiver fails on subsequent legitimate data.
Methodology: Transfer-based Black-Box Poisoning
In a realistic scenario, an attacker is unlikely to know the exact architecture or weights of the target deep receiver. To overcome this, the authors propose a transfer-based black-box attack pipeline consisting of four primary stages:
- Data Collection: The attacker monitors the communication channel to collect legitimate pilot data and channel observations.
- Surrogate Model Training: The attacker trains a generic "surrogate" DNN model that mimics the behavior of a standard receiver. To ensure the attack is robust, this model is trained across a wide range of signal-to-noise ratios (SNRs).
- Perturbation Generation: Using the surrogate model, the attacker employs Projected Gradient Descent (PGD) to find the optimal adversarial perturbation ($\delta$). This perturbation is designed to maximize the loss of the surrogate model while remaining small enough to be difficult to detect.
- The Poisoning Phase: The attacker transmits the adversarial perturbation into the air. The legitimate receiver picks up these "poisoned" pilots and uses them for its online update, effectively "deactivating" its own decoding capabilities by learning a corrupted representation of the channel.
Experimental Evaluation and Channel Environments
The effectiveness of the attack was tested across four distinct channel environments to ensure generalizability:
- Synthetic Linear and Nonlinear Channels: Baselines for standard propagation and hardware-induced distortions (e.g., power amplifier nonlinearities).
- Static Channel: A control environment to measure the baseline impact of poisoning.
- COST 2100 Channel: A realistic, geometry-based stochastic channel model used for high-fidelity MIMO simulations.
Key Insights: The Meta-Learning Paradox
A significant finding of this research is the Meta-Learning Paradox. Receivers like Meta-DeepSIC, which are specifically designed to be highly sensitive and fast-adapting to new environments, proved to be more vulnerable to poisoning than standard architectures. Because meta-learning models are optimized to learn from very few samples, they are more likely to "overfit" to the malicious patterns injected by the attacker, leading to a faster and more complete breakdown of communication.
Furthermore, the attack demonstrated high architecture agnosticism. Perturbations generated using a simple feedforward surrogate model were highly effective at poisoning complex, model-based receivers like DeepSIC, suggesting that the vulnerability is inherent to the way DNNs process wireless signals.
Results and Performance Impact
The poisoning attack consistently degraded the Symbol Error Rate (SER) across all tested scenarios. In nonlinear channels at 15 dB SNR, the attack caused SER deteriorations of up to 3.09 dB for Meta-DeepSIC and 2.74 dB for standard DeepSIC. In rapidly changing channels, poisoned receivers lost their ability to track the channel state, resulting in a total failure to decode information symbols after only a few dozen data blocks.
While increasing the number of pilots (e.g., from 200 to 1000) was found to mitigate some of the attack's impact, this solution is inefficient as it significantly reduces spectral efficiency, leaving less bandwidth for actual data transmission.
Conclusion and Future Directions
This work highlights a critical security gap in the deployment of adaptive deep receivers. The ability of an attacker to "blind" a receiver through pilot poisoning poses a serious threat to the reliability of next-generation MIMO systems. Future research must focus on developing poisoning purification techniques to sanitize pilot signals and designing robust architectures that can distinguish between legitimate channel shifts and malicious adversarial interference.
Summary updated