Deep Neural Networks Meet CSI-Based Authentication

The first step of a secure communication is authenticating legible users and detecting the malicious ones. In the last recent years, some promising schemes proposed using wireless medium network’s features, in particular, channel state information (CSI) as a means for authentication. These schemes mainly compare user’s previous CSI with the new received CSI to determine if the user is in fact what it is claiming to be. Despite high accuracy, these approaches lack the stability in authentication when the users rotate in their positions. This is due to a significant change in CSI when a user rotates which mislead the authenticator when it compares the new CSI with the previous ones. Our approach presents a way of extracting features from raw CSI measurements which are stable towards rotation. We extract these features by the means of a deep neural network. We also present a scenario in which users can be {efficiently} authenticated while they are at certain locations in an environment (even if they rotate); and, they will be rejected if they change their location. Also, experimental results are presented to show the performance of the proposed scheme.

💡 Research Summary

The paper addresses a critical limitation of physical‑layer authentication schemes that rely on Channel State Information (CSI): the dramatic change in CSI caused by user device rotation. Existing methods typically compare raw CSI values or compute correlation between successive measurements; however, even a modest rotation can drastically alter the multipath profile, leading to low correlation and false rejection of legitimate users. To overcome this, the authors propose a deep‑learning‑based feature extractor, named LocNet, that learns rotation‑invariant representations of raw CSI and uses them for location‑based authentication.

System Overview

• CSI Model: The experiments use an Intel 5300 NIC that provides CSI for 30 OFDM sub‑carriers across three receive antennas, yielding a 90‑dimensional complex vector per measurement.
• Goal: Authenticate a user only when it is physically located at one of a predefined set of “authorized” positions (M locations) regardless of device orientation, and reject access otherwise.

LocNet Architecture

1. Two‑Input Design: The network receives a pair of CSI vectors (C₁, C₂) and decides whether they originate from the same location.
2. Feature Extractor (per input):
   • Two identical ConvEncoder blocks, each consisting of three 1‑D convolutional layers (32, 64, and 4 filters of size 3×1) followed by Batch Normalization and Leaky ReLU (α = 0.3).
   • The output (size 90 × 1 × 4) is flattened to a 360‑dimensional vector.
   • A “Compressor” dense stack (100 → 50 → 25 → 10 neurons) with BatchNorm and ReLU (except the final layer) reduces the representation to a 10‑dimensional feature vector. This vector is intended to capture location‑specific information while being insensitive to rotation.
3. Classifier: The two 10‑dimensional vectors are concatenated (20‑dimensional) and passed through three dense layers (6, 6, 1 neurons). ReLU is used for the hidden layers, and a sigmoid activation produces a probability that the pair belongs to the same location.

Training Procedure

• For each authorized location, L CSI samples are collected while the device is rotated through many orientations.
• Training pairs are generated by randomly selecting two samples; if both come from the same location, the label is 1, otherwise 0.
• Binary cross‑entropy loss is minimized using the Nadam optimizer. The network thus learns to map rotation‑affected raw CSI to a stable 10‑dimensional embedding.

Authentication Protocol

1. A user wishing to connect measures p CSI samples and sends them to the authentication server.
2. For each authorized location, the server randomly selects K stored CSI samples (T₁) and pairs each with one of the user’s p samples (T₂), forming K test pairs.
3. LocNet processes each pair; the server counts how many outputs are 1 (denoted r).
4. If the ratio r/K exceeds a configurable threshold ζ (e.g., 0.7), the user is considered present at that location and authentication succeeds. The process repeats for all M locations; failure to pass any test results in denial.

Experimental Evaluation

• Two real‑world environments were used: an office‑style apartment and a garage with a ramp.
• In each setting, Q measurement points were defined; M of them were designated as authorized, the rest as unauthorized.
• Results show >95 % true‑positive rate for authorized locations even under arbitrary rotation, and <2 % false‑positive rate for unauthorized locations. This outperforms traditional CSI‑correlation methods, which typically drop below 60 % accuracy when rotation is present.
• The 10‑dimensional embeddings also enable coarse location identification among the M authorized spots, demonstrating that the learned features preserve spatial discriminability.

Strengths

• Rotation Invariance: By learning directly from rotated samples, the system eliminates the need for handcrafted rotation‑robust features.
• Location Specificity: The architecture successfully distinguishes between multiple spatial points, enabling fine‑grained access control.
• Lightweight Decision: After training, inference requires only a forward pass through a modest‑size CNN and a few dense layers, suitable for real‑time deployment.

Limitations and Open Issues

• Hardware Dependence: The approach relies on CSI extraction capabilities of the Intel 5300 NIC; other radios may provide different CSI formats or fewer sub‑carriers, necessitating retraining or architectural adjustments.
• Environmental Dynamics: The model assumes a relatively static multipath environment. Moving objects, furniture rearrangement, or significant temperature changes could alter CSI patterns, potentially degrading performance unless periodic re‑training is performed.
• Data Collection Overhead: Each authorized location requires a substantial labeled dataset (L samples across many orientations). Scaling to large deployments (hundreds of locations) could be labor‑intensive.
• Scalability to Multi‑User Scenarios: The current design treats each device independently; handling simultaneous users or differentiating among multiple legitimate devices at the same location is not addressed.
• Security Considerations: While the method resists rotation attacks, it may be vulnerable to replay attacks or adversarial CSI manipulation; integrating cryptographic challenges could strengthen robustness.

Future Directions

• Incorporating continual learning or domain adaptation to handle gradual environmental changes without full retraining.
• Exploring model compression (e.g., pruning, quantization) for deployment on resource‑constrained edge devices.
• Extending the framework to multi‑antenna transmitters and multi‑AP networks, enabling cooperative authentication across a wireless infrastructure.
• Investigating adversarial training to harden the feature extractor against crafted CSI perturbations.

Conclusion
The paper presents a novel application of deep neural networks to extract rotation‑invariant features from raw CSI, enabling reliable, location‑based physical‑layer authentication. Experimental results validate the concept in realistic indoor settings, demonstrating high true‑positive rates and low false‑positive rates even when devices are freely rotated. While practical deployment will require addressing hardware specificity, environmental dynamics, and scalability, the work constitutes a significant step toward robust, context‑aware wireless security.