Binary Filtering Overview

• Binary filtering is a suite of methods that reduces information to binary states for selection, denoising, feature screening, and secure processing across various applications.
• Techniques such as binary stochastic filtering use Bernoulli-based gating with sparsity penalties to efficiently prune neural network features while maintaining accuracy.
• Implementations in image processing, metric feature selection, Bayesian inference, and set membership leverage adaptive and probabilistic designs, yet require careful tradeoffs for noise and computational complexity.

Binary filtering encompasses a suite of techniques, algorithms, and data structures that operate on binary data—either for information filtration, signal/image denoising, feature selection, cryptographic sequence processing, or set membership queries—by means of Boolean, probabilistic, or threshold-based operations. The unifying characteristic is the explicit or implicit reduction of information or objects to binary states, and ensuing processing steps that apply binary-valued criteria either for selection, suppression, or transformation. Applications range from neural network model pruning, metric feature screening, and image morphology to hash-based set filters and cryptographically secure keystream generation.

1. Binary Stochastic Filtering for Machine Learning

Binary stochastic filtering (BSF) is a differentiable, probabilistic method for selecting informative subsets of features, hidden units, or regions within neural architectures. A BSF layer is parameterized by per-input Bernoulli gates: for each input $x_j$, a binary mask $b_j \sim \mathrm{Bernoulli}(p_j)$ determines passing or suppression. The probabilities $p_j$ are trained jointly with model parameters under an $L^1$ penalty to induce sparsity (Trelin et al., 2019, Trelin et al., 2020):

$$\mathrm{bsf}(x)_j = x_j b_j,\qquad b_j \sim \mathrm{Bernoulli}(p_j)$$

$$L_{\mathrm{total}}(\theta, p) = \mathbb{E}_{b}[L_{\mathrm{task}}(f(\mathrm{bsf}(x), \theta), y^*)] + \lambda \sum_j p_j$$

Straight-through or reparameterized gradient estimators (e.g., with Gumbel-Softmax relaxation) are used to propagate learning signals through the stochastic gating. At inference, binary filtering is fixed by thresholding $p_j$. This approach demonstrates effective feature and neuron pruning with minimal computational overhead and negligible accuracy trade-off, supporting both fully connected and convolutional variants. Empirical benchmarks show multi-fold parameter reduction, competitive or superior to mutual information, decision tree, $L^1$-SVM, and wrapper-based methods (Trelin et al., 2020).

2. Morphological and Stack-Based Binary Filtering in Image Processing

Binary filtering is fundamental to mathematical morphology, where binary images are processed using set-theoretic erosions, dilations, openings, and closings defined over graphs or higher-order hypergraphs (Prakash et al., 2015):

• Graph-based: For a binary set $X\subset V$ of foreground pixels, dilation $\delta_G(X)$ and erosion $b_j \sim \mathrm{Bernoulli}(p_j)$0 are given by binary set operations over immediate neighborhoods, composing to classic morphological filters.
• Hypergraph-based: Morphology operators are extended to hypergraphs $b_j \sim \mathrm{Bernoulli}(p_j)$1, using four adjoint binary maps to realize vertex- and hyperedge-level erosions/dilations and associated openings/closings.

Alternating Sequential Filters (ASF) recursively compose these operators to robustly remove noise (e.g., salt-and-pepper) while preserving shape. Experimental results on binarized images show hypergraph-based ASF dramatically reduces mean square error at low to moderate noise levels, outperforming both standard graph-based ASF and the median filter under these regimes (Prakash et al., 2015).

Stack filtering provides another nonlinear framework for grey-scale and binary image denoising (Buemi et al., 2012). The image is threshold-decomposed into binary slices, each filtered by a learned positive Boolean function applied to a local neighborhood. Adaptive design allows data-driven optimization of the Boolean rule against a noiseless reference. Stack filters achieve superior edge preservation—measured by the correlation of Laplacians and universal quality metrics (Q-index)—and improved classification accuracy compared to classical despeckling filters.

3. Binary Filtering in Feature Selection for Metric and Object Data

Binary filtering paradigms extend to high-dimensional, non-vectorial object spaces via statistic-based screening. The Metric Kolmogorov Filter (MK-Filter) generalizes Kolmogorov-Smirnov divergence to metric spaces for model-free screening of random objects (e.g., distributions, SPD matrices) for binary classification (He et al., 2024). Each feature is scored by a symmetrized metric Kolmogorov distance between class-conditional distributions:

$b_j \sim \mathrm{Bernoulli}(p_j)$2

A two-sample splitting and thresholding scheme guarantees both the sure-screening property (retention of all truly informative features) and asymptotic control of the false discovery rate. Empirical evaluation with Wasserstein and Log–Cholesky metrics demonstrates superior performance to Euclidean and classical nonparametric screeners, enabling reliable pre-filtering over tens of thousands of distributional or connectivity-matrix objects.

4. Binary Bayesian Filtering for Spatial and Sequence Inference

Binary filtering schemes underlie Bayesian spatial mapping when quantities of interest (e.g., LoS/NLoS link states) are binary and must be inferred from noisy measurements. The binary Bayesian filter recursively estimates the posterior log-odds for each spatial location conditioned on observed data, using an initial empirical prior and measurement likelihoods (Yang et al., 2024):

$b_j \sim \mathrm{Bernoulli}(p_j)$3

Spatial correlation is incorporated via radial and angular kernels, enabling the efficient and robust estimation of entire 2D binary maps (e.g., link state maps in urban UAV cellular coverage) under sparse measurement regimes. Empirical results confirm strong gains over KNN interpolation and non-updating benchmarks, especially with limited data.

5. Binary Filtering Functions in Cryptographic LFSR Generators

Binary filtering in cryptography refers to Boolean-valued functions applied to the binary states of Linear Feedback Shift Registers (LFSRs), used in pseudorandom keystream generators. Given an LFSR of degree $b_j \sim \mathrm{Bernoulli}(p_j)$4 and a nonlinear filtering function $b_j \sim \mathrm{Bernoulli}(p_j)$5 (expressed in algebraic normal form), binary filtering produces keystreams of high complexity (Fúster-Sabater et al., 2022). A key structural property is that, for any LFSR-filter pair, it is possible to construct a weakly equivalent filter on a reciprocal LFSR yielding the same output sequence but potentially lower algebraic complexity:

• The output sequence is invariant under the transformation to a reciprocal LFSR and corresponding nonlinear filter $b_j \sim \mathrm{Bernoulli}(p_j)$6, which can be algorithmically derived by transplanting the cyclotomic coset representation of $b_j \sim \mathrm{Bernoulli}(p_j)$7.
• This weak equivalence is critical for cryptanalysis, since security must consider the strongest (i.e., lowest-order) equivalent filter across all LFSRs of the same length.

The ability to explicitly construct weak equivalents resolves a longstanding open problem and facilitates rigorous cryptographic evaluation.

6. Binary Filters as Set Membership Data Structures

Binary filters encompass probabilistic data structures (Bloom, XOR, Ribbon filters) for resource-efficient set membership queries, widely used in databases, networking, and storage systems (Dillinger et al., 2021). Such filters encode membership information via bitwise operations, permitting false positives with well-characterized rates:

Filter Type	Space Overhead (%)	FP Rate Target Range	Query/Construction Speed
Bloom	≈44	$b_j \sim \mathrm{Bernoulli}(p_j)$8	Fastest query (α>40%)
XOR	≈22 (for $b_j \sim \mathrm{Bernoulli}(p_j)$9)	$p_j$0	Moderate
Ribbon (Standard)	2–20	$p_j$1 to $p_j$2	Fast construction, slower query
Ribbon (Balanced)	<1	$p_j$3	High construction cost, low query time

Ribbon filters are constructed by solving a banded Boolean linear system, permitting exceptionally low space overhead (as low as <1%) for practical $p_j$4. Variants such as Homogeneous Ribbon and Balanced Ribbon further optimize for simplicity or reduced overhead, respectively. This class of binary filtering data structures narrows the gap to the information-theoretic limit for space-per-key versus false positive rate, greatly benefitting static set applications with stringent space or construction constraints.

7. Cross-Domain Synthesis and Limitations

Despite highly diverse instantiations, binary filtering techniques share several recurring themes: reliance on Boolean operations, the use of probabilistic or adaptive selection, and potential for large-scale, computationally efficient information reduction. Common technical ingredients include regularization-based sparsification, structured Boolean function design, and Bayesian updating under binary hypotheses.

Limitations and considerations vary by domain. In neural BSF, independent gate assumption may impair group feature selection in highly correlated input regimes (Trelin et al., 2020). Hypergraph-based image filters excel at moderate noise but may over-smooth fine detail under high noise (Prakash et al., 2015). Metric-space feature screening is conditioned on the choice of the underlying metric and computational scalability (He et al., 2024). Set-membership binary filters must balance false positive rates against the complexity of linear-system solvers (Dillinger et al., 2021). In cryptographic sequence design, weakness in a reciprocal binary filter may undermine overall system security (Fúster-Sabater et al., 2022).

Altogether, binary filtering provides a foundational abstraction for discrete selection, denoising, summarization, and inference, underpinned by advances in neural, morphological, statistical, data-structural, and cryptographic methodologies across contemporary computational research.