Data-Driven Algorithms for Robust or Selective CFAR Detection in Colored Gaussian Noise

Ensuring a constant false alarm rate (CFAR) is a key requirement in radar detection. Several model-based CFAR detectors for colored Gaussian noise have been proposed, aiming at striking a desirable trade-off between robust or selective behavior and limited performance loss under matched conditions. More recently, data-driven methods (SVM, KNN, neural networks) have been explored, but they lack theoretical CFAR guarantees unless applied to maximal invariant statistics. In this paper, we introduce two novel, low-complexity, learning-based CFAR detectors that leverage maximal invariant statistics as input features: (i) a lightweight neural net-work with residual encoder blocks, and (ii) a modified single-layer network with random features. Their performance is benchmarked against model-based detectors and existing machine learning approaches, revealing new trade-offs and improved detection performance while maintaining execution times comparable to those of classical model-based solutions.