Channel Prediction-Based Physical Layer Authentication under Consecutive Spoofing Attacks

1. Problem Statement

Context: Wireless networks are vulnerable to spoofing attacks where an adversary (Mallory) impersonates a legitimate user (Bob) to transmit malicious data to a receiver (Alice). Physical Layer Authentication (PLA) typically uses Channel State Information (CSI) as a unique fingerprint to verify identity.

The Challenge:

• Consecutive Spoofing: While existing PLA methods handle single-packet attacks well, they fail under consecutive spoofing attacks (where the attacker transmits a sustained stream of malicious packets).
• Channel Evolution: In mobile environments, the legitimate channel between Bob and Alice changes over time due to mobility and fading.
• The Failure Mode: Traditional PLA relies on comparing the current CSI against a reference CSI from the last legitimate transmission. During a long spoofing interval, the legitimate channel evolves significantly. When Bob resumes transmission, the accumulated channel decorrelation causes the system to falsely reject legitimate packets (high False Alarm Rate) because the reference is outdated.

2. Methodology

The authors propose a Channel Prediction-Based PLA Framework that integrates a Transformer-based neural network with a correlation-based authentication mechanism. The system operates in two stages:

A. System Architecture

• Input: Historical CSI measurements from the legitimate user (Bob).
• Core Component: A Transformer model (Encoder-Decoder structure) designed to capture long-range temporal correlations in wireless channels.
• Output: Predicted future CSI measurements for the legitimate channel during the spoofing interval.

B. Training Stage

• Data: The model is trained on sequences of legitimate CSI.
• Task: Given $N_p$ historical CSI measurements, the Transformer predicts the next $N_f$ future CSI measurements.
• Loss Function: The model minimizes the Normalized Mean Squared Error (NMSE) between predicted and ground-truth CSI.
• Structure:
  • Encoder: Extracts temporal features from historical data using multi-head self-attention.
  • Decoder: Predicts future steps using masked self-attention (to prevent looking ahead) and cross-attention with the encoder output.
  • Position Selection: Only the last $N_f$ outputs of the decoder are used as the predicted CSI sequence.

C. Test Stage (Iterative Authentication)

The system operates iteratively to handle sustained attacks:

1. Prediction: The trained Transformer predicts the next $N_f$ legitimate CSI measurements based on the current input history.
2. Authentication: For each predicted step $j$, the system calculates the Pearson correlation between the predicted CSI and the observed incoming CSI.
   • If correlation $\ge$ threshold $\epsilon$: The packet is Legitimate.
   • If correlation $<$ threshold $\epsilon$: The packet is Malicious (Spoofed).
3. Adaptive Input Update (Key Innovation):
   • If a packet is Legitimate: The observed CSI is added to the input sequence for the next iteration.
   • If a packet is Malicious: The predicted CSI (not the observed malicious one) is added to the input sequence.
   • Significance: This prevents the "poisoning" of the prediction model. By feeding the predicted legitimate channel back into the model during an attack, the system maintains a reliable trajectory of the true channel evolution, ensuring that when the legitimate user returns, the system is ready to authenticate them correctly.

3. Key Contributions

1. Novel Framework: First to explicitly address consecutive spoofing attacks in PLA by predicting the evolution of the legitimate channel during the attack interval, rather than relying on a static reference.
2. Transformer Integration: Utilizes the Transformer architecture to model long-range temporal dependencies in wireless channels, overcoming the limitations of previous RNN/LSTM-based approaches.
3. Robust Input Update Strategy: Introduces an adaptive mechanism where the model input is updated with predicted CSI during attacks. This ensures the model remains robust against adversarial data injection and prevents error accumulation from corrupting the channel state estimation.
4. Comprehensive Evaluation: Validates the approach under Rayleigh fading channels with varying attack lengths, terminal speeds, and SNR levels.

4. Simulation Results

The authors evaluated the scheme using IEEE 802.11n OFDM parameters (20 MHz bandwidth, 64 subcarriers) under Rayleigh fading.

• Prediction Accuracy:
  • The Transformer achieved low Normalized Mean Squared Error (NMSE).
  • A prediction length ($N_f$) of 5 was found to offer the best trade-off between prediction accuracy per step and the accumulation of errors over multiple iterations.
• Authentication Performance:
  • Benchmark: A conventional Pearson correlation method (using the last legitimate packet as reference) performed well for short attacks ($N_a \le 3$) but degraded rapidly as the attack length increased due to channel decorrelation.
  • Proposed Method: Maintained high authentication accuracy even under extended spoofing attacks (large $N_a$).
  • Result: The proposed scheme significantly outperformed the benchmark in terms of sequence-level accuracy, successfully distinguishing between natural channel variations and malicious spoofing.

5. Significance

This research addresses a critical gap in Physical Layer Security: the vulnerability of existing PLA systems to sustained attacks in mobile environments.

• Practical Impact: It enables secure communication in dynamic IoT and vehicular networks where attackers may attempt to hijack connections for extended periods.
• Theoretical Advance: It demonstrates that deep learning models (Transformers), when coupled with adaptive feedback loops, can effectively model and predict physical channel dynamics, turning a vulnerability (channel variation) into a security asset.
• Resilience: The adaptive update strategy ensures that the authentication system does not "forget" the legitimate channel characteristics even when the legitimate user is silent for a long duration.