DeepAlloc: CNN-Based Approach to Efficient Spectrum Allocation in Shared Spectrum Systems

TL;DR

This work tackles efficient spectrum allocation in shared spectrum systems under incomplete primary-user information by learning the spectrum allocation function with CNNs. It introduces a CNN-based framework (DeepAlloc, with a shallow baseline SH-Alloc) that represents inputs as image-like sheets capturing PU/SS/SU parameters, and uses pre-training on synthetic propagation data to boost field performance. The approach includes techniques to minimize false positives, account for multipath fading, generate synthetic samples, and extend to multiple SUs with NN and RNN architectures. Across large-scale simulations and a small outdoor testbed, DeepAlloc achieves up to about 60% improvement over prior methods and attains near-optimal power allocations (3–4 dB away from the optimum), demonstrating meaningful gains in spectrum utilization while maintaining protections for PU receivers.

Abstract

Shared spectrum systems facilitate spectrum allocation to unlicensed users without harming the licensed users; they offer great promise in optimizing spectrum utility, but their management (in particular, efficient spectrum allocation to unlicensed users) is challenging. A significant shortcoming of current allocation methods is that they are either done very conservatively to ensure correctness, or are based on imperfect propagation models and/or spectrum sensing with poor spatial granularity. This leads to poor spectrum utilization, the fundamental objective of shared spectrum systems. To allocate spectrum near-optimally to secondary users in general scenarios, we fundamentally need to have knowledge of the signal path-loss function. In practice, however, even the best known path-loss models have unsatisfactory accuracy, and conducting extensive surveys to gather path-loss values is infeasible. To circumvent this challenge, we propose to learn the spectrum allocation function directly using supervised learning techniques. We particularly address the scenarios when the primary users' information may not be available; for such settings, we make use of a crowdsourced sensing architecture and use the spectrum sensor readings as features. We develop an efficient CNN-based approach (called DeepAlloc) and address various challenges that arise in its application to the learning the spectrum allocation function. Via extensive large-scale simulation and a small testbed, we demonstrate the effectiveness of our developed techniques; in particular, we observe that our approach improves the accuracy of standard learning techniques and prior work by up to 60%.

Figures (16)

• Figure 1: Eqn. (\ref{['eq:maxpower']}) Illustration. The optimal power that can be allocated to an SU is such that, at each PUR (a PU's receiver) the signal-to-noise ratio is more than the desired ratio, $\beta$. Above, $R_{j}$ is a certain PUR, $I_j$ is the total interference at $R_{j}$ from other PUs, $s_{j}$ is the signal strength received at $R_{j}$ from its PU, and $\Pi \cdot \rho(l\xspace, l_j)$ is the interference due to the SU at $R_{j}$ where $l\xspace$ and $l_j$ are their respective locations.
• Figure 2: Image representation of a sample for input to the DeepAlloc model. Here, there are four PUs (2 in the first sheet, and 1 in the other two sheets) and one SU (in the fourth sheet). Each sheet for the PUs corresponds to a range of PU's transmit power; e.g., PUs whose transmit power is in the range -10 to 0 dBm are placed in the first sheet. When representing SSs in SS-Setting, we place SSs in the sheets based on their locations.
• Figure 3: CNN-based Spectrum Allocation System. For the CNN Model component, SH-Alloc uses a shallow CNN model without any pre-training, while DeepAlloc uses a deep pre-trained CNN model obtained from Fig. \ref{['fig:deep_figure']}. In both cases, after any pre-training, the (field) training samples are gathered, converted into images, and then used to (further) train the CNN model. The learned model is then used to allocate spectrum to requesting SUs.
• Figure 4: DeepAlloc Pre-Training Process. Here, we generate a large number ($\approx$ 1M) of images assuming a log-normal propagation model based on an estimate path-loss exponent, and pre-train a deep CNN model (we used ResNet) using these generated images. The pre-trained model is then used in Fig. \ref{['fig:cnn_system']} for further training using a smaller number of field training samples.
• Figure 5: (a) Spectrum allocation function, when there is a single PU and a single SU, due to path loss with shadowing and multi-path fading effects. Note that the path-loss function has a similar trend. The red plot is a conservative spectrum allocation, based on a similar conservative path-loss function. (b) Modifying labels of training samples to drive the model towards learning a simpler and more conservative spectrum allocation function.