BRISC: A Dataset of Channel Measurements at 5 GHz With a Reflective Intelligent Surface

1. Problem Statement

The evolution toward 6G wireless systems requires active control of the propagation environment, primarily through Reconfigurable Intelligent Surfaces (RIS). While RIS technology promises enhanced flexibility, reliability, and security, its development is hindered by a critical lack of real-world experimental data.

• Data Scarcity: Existing datasets are predominantly simulated or limited to specific, narrow configurations (e.g., anechoic chamber measurements with only forward transmission coefficients).
• Research Gap: There is a shortage of comprehensive Channel State Information (CSI) datasets that capture the complex interplay between RIS configurations, multipath propagation, and transmitter/receiver geometry in realistic indoor environments. This scarcity prevents the reproducible validation of channel models, learning algorithms, and security protocols designed for RIS-assisted communications.

2. Methodology

The authors conducted a large-scale measurement campaign to create the BRISC (Broadband Reconfigurable Intelligent Surface Channel) dataset.

Experimental Setup

• Environment: A controlled indoor environment with RF absorbers to block Line-of-Sight (LOS) between the Transmitter (TX) and Receiver (RX), and to isolate RIS reflections from wall reflections.
• Hardware:
  • RIS: A 16×16 array (256 elements) of binary-state unit cells on a PCB. Each element uses a CMOS SPDT switch to toggle between two reflection states (ON/OFF), providing 1-bit control.
  • Transceiver: Two Ettus USRP X310 Software Defined Radios (SDRs) synchronized via an OctoClock (10 MHz reference + PPS) to ensure phase coherence.
  • Antennas: TX uses a directive log-periodic antenna; RX uses both a log-periodic and a dipole antenna to capture different multipath characteristics.
• Parameters:
  • Frequency: 5.53 GHz (center of the Wi-Fi band where the RIS exhibits maximum phase variation).
  • Bandwidth: 80 MHz (IEEE 802.11ac VHT frames).
  • Configurations: 9 distinct TX positions, with the RX and RIS fixed.

Data Acquisition Strategy

To manage the complexity of the $2^{256}$ configuration space, the authors employed a structured exploration strategy:

1. Macro-block Grouping: Elements were grouped into blocks (e.g., 4 blocks of 8×8, then 9 blocks in a 3×3 grid) where all elements in a block share the same state.
2. Random Configurations: Fully random element-wise ON/OFF assignments.
3. Procedure: For each TX position, the system cycled through RIS configurations with a settling time (~0.3s), transmitted ~60 frames, and included silence intervals to separate configurations.
4. Volume: Approximately 10,000 distinct RIS configurations were explored per TX position, resulting in a massive dataset of complex baseband IQ samples and extracted CSI.

3. Key Contributions

1. The BRISC Dataset: A publicly available, comprehensive dataset containing ~10,000 configurations per TX position across 9 locations, covering binary RIS states, varying block sizes, and dual-antenna reception.
2. Channel Model Validation: The paper validates state-of-the-art RIS channel models against empirical data. It compares:
   • Linear Models (LM/LMB): Mapping RIS states to channel response using Ridge regression.
   • Non-Linear Models (NLM): Using Neural Networks (NN) and Random Forests (RAF) for data-driven inference.
3. Physical-Layer Authentication (PLA) Use Case: A demonstration of how RIS-assisted channels can be used for security, specifically distinguishing between a legitimate user and an impersonator based on spatial signatures.

4. Results

• Model Performance:
  • Non-linear models (NN, RAF) generally outperformed linear models when trained on small datasets.
  • Linear models with bias (LMB) proved highly effective, particularly when training on macro-block configurations. Surprisingly, LMB achieved an NMSE of -20.57 dB on unseen data, performing comparably to or better than Neural Networks in specific reconstruction tasks.
  • Key Insight: Changing blocks of elements creates higher channel diversity than random single-element changes, aiding linear model fitting.
• Physical-Layer Authentication (PLA):
  • The RIS creates a unique, location-dependent "spatial signature" that is difficult for an adversary to replicate.
  • Detection Performance: Using Detection Error Tradeoff (DET) curves, the study showed that even with a minimal number of subcarriers (N=1 or N=2), strong discrimination between a legitimate user (Alice) and an impersonator (Eve) is possible.
  • Efficiency: Increasing subcarriers beyond 2 yielded diminishing returns, suggesting that the spatial information provided by the RIS is highly concentrated.
  • Constraint: The authentication process was optimized to select RIS configurations that maintain at least 90% of the maximum communication capacity, ensuring security does not compromise throughput.

5. Significance

• Benchmarking: BRISC serves as a robust, reproducible benchmark for the research community to test signal processing, machine learning, and security algorithms for 6G RIS systems.
• Model Calibration: The dataset provides the empirical evidence needed to calibrate theoretical channel models, revealing that simple linear models with bias can be surprisingly effective, potentially reducing the computational cost of RIS control algorithms.
• Security Advancement: The work demonstrates that RIS can be leveraged not just for coverage, but as a critical enabler for Physical-Layer Security, offering a mechanism to authenticate users based on the unique physics of the propagation environment without relying solely on cryptographic keys.
• Realism: By capturing data in a realistic indoor setting with LOS blocked and multipath present, the dataset bridges the gap between theoretical simulations and deployable RIS technologies.

The dataset and analysis scripts are made available by the authors to facilitate further research in RIS-assisted communications.